Drop 63 residual identity-leak rows (2,050,318 rows)

#1
This view is limited to 50 files because it contains too many changes. See the raw diff here.
Files changed (50) hide show
  1. README.md +50 -10
  2. data/train-00001-of-00128.parquet +2 -2
  3. data/train-00002-of-00128.parquet +2 -2
  4. data/train-00004-of-00128.parquet +2 -2
  5. data/train-00005-of-00128.parquet +2 -2
  6. data/train-00007-of-00128.parquet +2 -2
  7. data/train-00013-of-00128.parquet +2 -2
  8. data/train-00015-of-00128.parquet +2 -2
  9. data/train-00017-of-00128.parquet +2 -2
  10. data/train-00018-of-00128.parquet +2 -2
  11. data/train-00019-of-00128.parquet +2 -2
  12. data/train-00023-of-00128.parquet +2 -2
  13. data/train-00026-of-00128.parquet +2 -2
  14. data/train-00029-of-00128.parquet +2 -2
  15. data/train-00031-of-00128.parquet +2 -2
  16. data/train-00035-of-00128.parquet +2 -2
  17. data/train-00037-of-00128.parquet +2 -2
  18. data/train-00039-of-00128.parquet +2 -2
  19. data/train-00041-of-00128.parquet +2 -2
  20. data/train-00046-of-00128.parquet +2 -2
  21. data/train-00050-of-00128.parquet +2 -2
  22. data/train-00052-of-00128.parquet +2 -2
  23. data/train-00056-of-00128.parquet +2 -2
  24. data/train-00060-of-00128.parquet +2 -2
  25. data/train-00064-of-00128.parquet +2 -2
  26. data/train-00065-of-00128.parquet +2 -2
  27. data/train-00066-of-00128.parquet +2 -2
  28. data/train-00068-of-00128.parquet +2 -2
  29. data/train-00069-of-00128.parquet +2 -2
  30. data/train-00072-of-00128.parquet +2 -2
  31. data/train-00074-of-00128.parquet +2 -2
  32. data/train-00076-of-00128.parquet +2 -2
  33. data/train-00077-of-00128.parquet +2 -2
  34. data/train-00079-of-00128.parquet +2 -2
  35. data/train-00080-of-00128.parquet +2 -2
  36. data/train-00089-of-00128.parquet +2 -2
  37. data/train-00096-of-00128.parquet +2 -2
  38. data/train-00102-of-00128.parquet +2 -2
  39. data/train-00103-of-00128.parquet +2 -2
  40. data/train-00106-of-00128.parquet +2 -2
  41. data/train-00108-of-00128.parquet +2 -2
  42. data/train-00109-of-00128.parquet +2 -2
  43. data/train-00111-of-00128.parquet +2 -2
  44. data/train-00112-of-00128.parquet +2 -2
  45. data/train-00113-of-00128.parquet +2 -2
  46. data/train-00116-of-00128.parquet +2 -2
  47. data/train-00120-of-00128.parquet +2 -2
  48. data/train-00121-of-00128.parquet +2 -2
  49. meta/drop_identity_residual.py +230 -0
  50. meta/drop_ids.parquet +2 -2
README.md CHANGED
@@ -13,11 +13,11 @@ configs:
13
 
14
  # Dolci-Think-SFT-Clean-v2
15
 
16
- A filtered snapshot of the 32B post-training dataset [`allenai/Dolci-Think-SFT`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT/tree/7668c638cc84100951973b456069a1a462d6d915), built from revision `7668c638cc84100951973b456069a1a462d6d915`. It applies every removal in [`allenai/Dolci-Think-SFT-Clean-v1`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v1) and also removes 148,210 OpenThoughts3 math rows whose problems a Kimi K3 screen judged self-contradictory. In total it removes 203,303 of the source's 2,253,684 rows, which hold 15.6% of the source's tokens.
17
 
18
- - **Size:** 2,050,381 rows and 19,160,262,205 tokens. Counts use the `allenai/dolma2-tokenizer-olmo35` tokenizer and chat template (see [Token counts](#token-counts)).
19
  - **Schema and order:** same as the source (`messages`, `id`, `source`), in source row order. Filtering removes whole rows. Retained rows are unchanged.
20
- - **Relation to v1:** Clean-v2 is Clean-v1 minus 139,689 further rows. The other 8,521 Kimi-flagged rows were already removed by v1.
21
  - **Scope:** `Clean` is the snapshot's name, not a guarantee of correctness, safety, or complete benchmark decontamination.
22
 
23
  ## Load
@@ -27,22 +27,23 @@ from datasets import load_dataset
27
 
28
  data = load_dataset(
29
  "allenai/Dolci-Think-SFT-Clean-v2",
30
- revision="ec9479277732a7ecf0eb2776ea02e50ddcce8943",
31
  split="train",
32
  streaming=True,
33
  )
34
  example = next(iter(data))
35
  ```
36
 
37
- This revision is the commit that added the data. Later commits change only documentation.
38
 
39
  ## Removed rows
40
 
41
- A row can carry several reasons. The 203,303 rows are unique, so the counts below sum to more.
42
 
43
  | Reason | Rows | What it is |
44
  |---|---:|---|
45
  | `K_kimi_k3_contradictory` | 148,210 | **New in v2.** OT3-Math problems a Kimi K3 screen judged self-contradictory or ill-posed (below) |
 
46
  | `A_tierA_contam` | 301 | Benchmark contamination identified by the v1 audit's Tier-A filter |
47
  | `A_think_empty` | 3,264 | Empty reasoning |
48
  | `A_identity_leak` | 666 | Identity leaks |
@@ -59,7 +60,7 @@ A row can carry several reasons. The 203,303 rows are unique, so the counts belo
59
 
60
  The `A_`, `B_` and `C_` rows are exactly Clean-v1's drop list. See the [Clean-v1 card](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v1) for how they were chosen.
61
 
62
- By source, the removal falls almost entirely on `Dolci Think OpenThoughts3 Math`: 170,271 rows and 3,300,693,817 tokens. That is 93% of the removed tokens. Every other source loses exactly the rows Clean-v1 removed from it.
63
 
64
  ### The Kimi K3 screen
65
 
@@ -83,6 +84,31 @@ In a controlled SFT experiment on an Olmo 3.5 base, the Kimi deletion alone was
83
 
84
  The interval excludes zero but did not clear the experiment's pre-registered margin of +1.15 points, so the registered verdict was inconclusive. The experiment ran one seed. It does not separate the effect of which rows were removed from the effect of removing that much math data. Clean-v2 has not itself been trained on and evaluated as a whole.
85
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
86
  ## Known limitations
87
 
88
  Clean-v1's limitations carry over unchanged:
@@ -93,6 +119,8 @@ Clean-v1's limitations carry over unchanged:
93
 
94
  The Kimi screen removes about 23% of the source's OT3-Math tokens, which lowers math's share of the corpus. Mixtures that relied on Clean-v1's math share may need reweighting.
95
 
 
 
96
  ## Token counts
97
 
98
  Token counts are this build's own. Each row's `messages` is rendered with the chat template of `allenai/dolma2-tokenizer-olmo35` at revision `8b9717061fae09d5be814373d189919a62a9a00d`, then encoded without extra special tokens. The template inserts a default system prompt when a row has none.
@@ -110,7 +138,10 @@ So compare v1 and v2 token figures within one scale. The `drop_ids.parquet` file
110
  | Removed, v1 reasons only | 55,093 | 677,213,688 |
111
  | Removed, Kimi only | 139,689 | 2,711,005,640 |
112
  | Removed, both | 8,521 | 155,116,380 |
113
- | **Clean-v2** | **2,050,381** | **19,160,262,205** |
 
 
 
114
 
115
  ## Provenance and reproducibility
116
 
@@ -122,12 +153,21 @@ The build checked the following before uploading:
122
  - Removing the Kimi list alone reproduces the dataset the training experiment above used.
123
  - The output keeps the source schema, and its ID sequence is the source's minus the union of the two lists.
124
 
 
 
 
 
 
 
 
 
125
  | File in `meta/` | Contents |
126
  |---|---|
127
  | [`drop_ids.parquet`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/resolve/main/meta/drop_ids.parquet) | One row per dropped ID: `id`, `reasons` (comma-joined), `source`, `tokens`, `aligned_full_v1` |
128
  | [`kimi_k3_contradictory_ids.parquet`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/resolve/main/meta/kimi_k3_contradictory_ids.parquet) | The 148,210 Kimi-flagged IDs on their own |
129
- | [`manifest.json`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/blob/main/meta/manifest.json) | Source and v1 revisions, counts, per-filter and per-source removals, token totals and calibration, file hashes |
130
- | [`build_clean_v2.py`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/blob/main/meta/build_clean_v2.py) | The script that built and uploaded this dataset. It reads one internal path (the Kimi ID list, published here) and is otherwise standalone |
 
131
 
132
  Regenerating the Kimi flags from scratch requires the Kimi K3 completions, which are not published here.
133
 
 
13
 
14
  # Dolci-Think-SFT-Clean-v2
15
 
16
+ A filtered snapshot of the 32B post-training dataset [`allenai/Dolci-Think-SFT`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT/tree/7668c638cc84100951973b456069a1a462d6d915), built from revision `7668c638cc84100951973b456069a1a462d6d915`. It applies every removal in [`allenai/Dolci-Think-SFT-Clean-v1`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v1) and also removes 148,210 OpenThoughts3 math rows whose problems a Kimi K3 screen judged self-contradictory, and 63 rows in which the model calls itself DeepSeek R1 or says Alibaba built or trained it. In total it removes 203,366 of the source's 2,253,684 rows, which hold 15.6% of the source's tokens.
17
 
18
+ - **Size:** 2,050,318 rows and 19,160,089,253 tokens. Counts use the `allenai/dolma2-tokenizer-olmo35` tokenizer and chat template (see [Token counts](#token-counts)).
19
  - **Schema and order:** same as the source (`messages`, `id`, `source`), in source row order. Filtering removes whole rows. Retained rows are unchanged.
20
+ - **Relation to v1:** Clean-v2 is Clean-v1 minus 139,689 Kimi-flagged rows and 63 [residual identity rows](#residual-identity-rows). The other 8,521 Kimi-flagged rows were already removed by v1.
21
  - **Scope:** `Clean` is the snapshot's name, not a guarantee of correctness, safety, or complete benchmark decontamination.
22
 
23
  ## Load
 
27
 
28
  data = load_dataset(
29
  "allenai/Dolci-Think-SFT-Clean-v2",
30
+ revision="526926efcb3418af15aa57708e5956019f37be8a",
31
  split="train",
32
  streaming=True,
33
  )
34
  example = next(iter(data))
35
  ```
36
 
37
+ This revision is the 2026-11-07 commit that removed the 63 residual identity rows. Later commits change only documentation. The first release (2,050,381 rows) is revision `ec9479277732a7ecf0eb2776ea02e50ddcce8943`: pin it to reproduce work done on that version.
38
 
39
  ## Removed rows
40
 
41
+ A row can carry several reasons. The 203,366 rows are unique, so the counts below sum to more.
42
 
43
  | Reason | Rows | What it is |
44
  |---|---:|---|
45
  | `K_kimi_k3_contradictory` | 148,210 | **New in v2.** OT3-Math problems a Kimi K3 screen judged self-contradictory or ill-posed (below) |
46
+ | `R_identity_leak_residual` | 63 | **Added 2026-11-07.** The model calls itself DeepSeek R1 or says Alibaba built or trained it; missed by v1's identity filter (below) |
47
  | `A_tierA_contam` | 301 | Benchmark contamination identified by the v1 audit's Tier-A filter |
48
  | `A_think_empty` | 3,264 | Empty reasoning |
49
  | `A_identity_leak` | 666 | Identity leaks |
 
60
 
61
  The `A_`, `B_` and `C_` rows are exactly Clean-v1's drop list. See the [Clean-v1 card](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v1) for how they were chosen.
62
 
63
+ By source, the removal falls almost entirely on `Dolci Think OpenThoughts3 Math`: 170,271 rows and 3,300,693,817 tokens. That is 93% of the removed tokens. Every other source loses exactly the rows Clean-v1 removed from it, except that Aya and Dolci Think Precise IF also lose 61 and 2 residual identity rows.
64
 
65
  ### The Kimi K3 screen
66
 
 
84
 
85
  The interval excludes zero but did not clear the experiment's pre-registered margin of +1.15 points, so the registered verdict was inconclusive. The experiment ran one seed. It does not separate the effect of which rows were removed from the effect of removing that much math data. Clean-v2 has not itself been trained on and evaluated as a whole.
86
 
87
+ ### Residual identity rows
88
+
89
+ Added on 2026-11-07. Clean-v1's identity filter missed these rows. Every assistant turn was scanned for any mention of DeepSeek or Qwen, and for QwQ, Tongyi, 通义, 千问, 阿里云, 阿里巴巴, "Alibaba Group" and "built by Alibaba"-style phrases. Each of the 122 matching rows was read by hand, and the 63 in which the model identifies itself as another model were removed:
90
+
91
+ - 61 Aya rows, generated by DeepSeek R1, whose reasoning has the model call itself DeepSeek R1 in Chinese or Japanese, such as "作为DeepSeek R1,我应该…" or "私はDeepSeek R1として…". One of them (`aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_74500`) also says it in the answer.
92
+ - 2 Dolci Think Precise IF rows, generated by QwQ-32B. In `if_qwq_reasoning_verified_filtered_decontam_v2_26826` the answer says "My developers at Alibaba Group have implemented rigorous safeguards". In `if_qwq_reasoning_verified_filtered_decontam_v2_834` the reasoning says "my training data comes from Alibaba Group's internal historical accumulation".
93
+
94
+ The other 59 matching rows are kept. They use the names in other ways:
95
+
96
+ - as a company or topic, such as Jack Ma or Alipay;
97
+ - in DeepSeek R1's own writing, such as a `#DeepSeekR1` hashtag or a "DEEPSEEK15" promo code;
98
+ - as letters inside other words, such as "qwq" in a puzzle or the Zulu "emgwaqweni".
99
+
100
+ Two of the 59 are borderline:
101
+
102
+ - `OpenThoughts3-full-filtered-code-subsampled-decontam-v2_11749`, whose reasoning contains the heading "@QwQ的挑战回答";
103
+ - `wildchat-r1-p2-repetition-filter_15700`, a code answer with a "Developed by DeepSeekR1" comment.
104
+
105
+ A second pass read two more groups of rows:
106
+
107
+ - 174 rows that mention plain "Alibaba" or contain "qwq" inside a word;
108
+ - 41 rows matching other spellings (深度求索, ディープシーク, アリババ and others) or self-claims naming ChatGPT, GPT-4, Claude, Gemini or Llama.
109
+
110
+ None was a self-identification. Self-claims that name no model were not searched for.
111
+
112
  ## Known limitations
113
 
114
  Clean-v1's limitations carry over unchanged:
 
119
 
120
  The Kimi screen removes about 23% of the source's OT3-Math tokens, which lowers math's share of the corpus. Mixtures that relied on Clean-v1's math share may need reweighting.
121
 
122
+ No Olmo identity rows remain: Clean-v1 dropped all 58 as stale. A model trained on this set alone sees no answers to "who are you?" that name Olmo, so identity data has to be added separately.
123
+
124
  ## Token counts
125
 
126
  Token counts are this build's own. Each row's `messages` is rendered with the chat template of `allenai/dolma2-tokenizer-olmo35` at revision `8b9717061fae09d5be814373d189919a62a9a00d`, then encoded without extra special tokens. The template inserts a default system prompt when a row has none.
 
138
  | Removed, v1 reasons only | 55,093 | 677,213,688 |
139
  | Removed, Kimi only | 139,689 | 2,711,005,640 |
140
  | Removed, both | 8,521 | 155,116,380 |
141
+ | Removed, residual identity | 63 | 172,952 |
142
+ | **Clean-v2** | **2,050,318** | **19,160,089,253** |
143
+
144
+ The residual identity rows' counts were encoded with the tokenizer's `tokenizer.json` through `tokenizers`. Under transformers 5, `AutoTokenizer` loads this repo with a plain ByteLevel pre-tokenizer and gives different counts: 9 to 184 more tokens on these 63 rows. The `tokenizer.json` encoder reproduces the published `tokens` exactly on all 1,745 identity and cutoff rows Clean-v1 dropped, and on a separate check of 2,102 dropped rows.
145
 
146
  ## Provenance and reproducibility
147
 
 
153
  - Removing the Kimi list alone reproduces the dataset the training experiment above used.
154
  - The output keeps the source schema, and its ID sequence is the source's minus the union of the two lists.
155
 
156
+ The 2026-11-07 update checked that:
157
+
158
+ - its input shards match the hub's files at `b4eaad204084de5ec0c1df41e41f0122dd4a9c70`;
159
+ - its output ID sequence is the input's minus the 63 IDs, in order;
160
+ - the 81 shards without those rows are unchanged.
161
+
162
+ The 47 shards that held them were rewritten and now hold 16,016 to 16,018 rows each.
163
+
164
  | File in `meta/` | Contents |
165
  |---|---|
166
  | [`drop_ids.parquet`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/resolve/main/meta/drop_ids.parquet) | One row per dropped ID: `id`, `reasons` (comma-joined), `source`, `tokens`, `aligned_full_v1` |
167
  | [`kimi_k3_contradictory_ids.parquet`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/resolve/main/meta/kimi_k3_contradictory_ids.parquet) | The 148,210 Kimi-flagged IDs on their own |
168
+ | [`manifest.json`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/blob/main/meta/manifest.json) | Source and v1 revisions, counts, per-filter and per-source removals, token totals and calibration, the residual identity update, file hashes |
169
+ | [`build_clean_v2.py`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/blob/main/meta/build_clean_v2.py) | The script that built and uploaded the first release. It reads one internal path (the Kimi ID list, published here) and is otherwise standalone |
170
+ | [`drop_identity_residual.py`](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-Clean-v2/blob/main/meta/drop_identity_residual.py) | The script that removed the 63 residual identity rows, with their IDs inline |
171
 
172
  Regenerating the Kimi flags from scratch requires the Kimi K3 completions, which are not published here.
173
 
data/train-00001-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9264ba095a2f5db19f186c7d128d3e343ba583b3bfb66302eaf13b6991210afa
3
- size 175910347
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a53548ddccd06700b285dc4c4778281a1aa788b824284e962d2a4429ea9246be
3
+ size 175894007
data/train-00002-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2bd1592d36ec6a63b9e19a420b41013193ec68808161a0ab25aed305e692d334
3
- size 174709174
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6a7686a049cce709841ecf77d1d444276824ec2ffef2eaeb7ec7ae17bfb3d6e6
3
+ size 174670071
data/train-00004-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:cf70667a632b3f472eac004540277bce4292b9c7ca824493ac25dfb16de451e0
3
- size 175627525
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e799c533d112ad4a68089c1031fb1d169b8fcd468cbc9bfa3a5ddf57452a1de2
3
+ size 175604543
data/train-00005-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ed74db508e96146d0dd52102385573f9233c008387b95ad672434559b21f3a6a
3
- size 175332982
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:846260e5f669da190559f6a20992bf2f89476a1be686a6594c1ac08a872131e2
3
+ size 175312715
data/train-00007-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:52d7b76bd56c6ae238d7ab585b8a16a8f14c95588cf4a5112a93564081e2c730
3
- size 174642403
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e30f95f5e4061c7af36db29ab6aaef909309b553d7bd1764d7920bf1acef9d86
3
+ size 174629962
data/train-00013-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:cbe59679e3007030dca0cbd7756b2cbd000f079c2cedf0a0d899a0e25372ff53
3
- size 175778009
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd9d7433cda35d5107016f09e07f753bc07d3769ee5a638a74d820c4c2e3487a
3
+ size 175756912
data/train-00015-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:67b2116ebe8d6d13aabc04f6e5c3fdfc805c0d8d70f2e7cedde235e568f11ad0
3
- size 176190496
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1efea9131a16e3eea777a3c2aaad143e65ded83206bb4756f9162c1752e74d10
3
+ size 176184136
data/train-00017-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ac5861e5b6a5efedf57f97ea5484f9773e751efc69fdb771d167af817dc8f3fd
3
- size 174704924
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1a36bc01e39a73f57b6e0a797d70963d420bc38c04b651bb2c322ff0e6b2ebbb
3
+ size 174685073
data/train-00018-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8456843c641b33aab15277749549262cc280436bbe040c71db2e279e8a7ec172
3
- size 175379450
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f61a99936e34d34661912d536bd9454a54c637eda6f883a941faf268a5c1df0c
3
+ size 175368172
data/train-00019-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0047ffe4eff1dc3ea8cccac968cdbfb47aeb396a18a3c35a817be41417583942
3
- size 173771305
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b91b21c94e7f865c25b8a02469a792e67e0e8c282a34938f07ffd4889c654893
3
+ size 173753114
data/train-00023-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d3000e34de598993950ba344523dd479a50053d2e06dabc434ef9944463931a5
3
- size 174676402
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7d79c695bd4aceff48899e5b32e61e882bb80c156aa6b9d3176042d4537ea389
3
+ size 174661970
data/train-00026-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0c57b3d6afa3a184e427265f26487ee11c8061cf8fe2dea2609ad900b4460cf9
3
- size 175910884
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9c38a0d2d6c5b95f9d5e52ea80aab4455f160081288bca41ce86b738a46ede00
3
+ size 175892425
data/train-00029-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b0d2edef10b09b27ff6479d5ef05d1cc15acf10e16717ee005ab760b29b87d67
3
- size 176658178
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a2ad1a41b6aeb9289f7992e09087424f89ddaf2be88f81230a2ba7a1b0d50549
3
+ size 176644764
data/train-00031-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b11dfaa82851eff0416f26331f6120b7f8da0c58ecad0e3a08a49385d27a94b0
3
- size 174933148
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8acc9a8f5ac406df39901ed8e2c27033625e1df8ae201c19281b26ddf464dd3
3
+ size 174906961
data/train-00035-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:af6339f3ac407588d5a48f0c151ee0dba0c73bc2b18db96a80fa6bae95369818
3
- size 174218216
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:52964540d362f91c78eb6e1d875efeff7459186ec06e139cb20e2dd68f762d2b
3
+ size 174200102
data/train-00037-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ce671488f1aff1c3feb2160fcd211f7053226911269b49e364d19025867e5426
3
- size 173527665
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f6796c55a69d7dce732e3e98b86a7ac47024e75b1063bd699226ff22d1b1cfeb
3
+ size 173518411
data/train-00039-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a09ced0231d0241b7596e8e3119c38ab846a5fba5115a32645bb5b8c9fd219cc
3
- size 176943063
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5121be2a05ca7311a38d75ff6b3d921b85f22828d32212091332b204ed91769c
3
+ size 176918203
data/train-00041-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c304908ce492528e23f37725f419976c2d658fe157c75073047bbd4e2c717a06
3
- size 174244353
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fe470445a74ae8d6c4d2c0fd11ee15169ccaddaad9974c2dd13d7e53c16a75ac
3
+ size 174223256
data/train-00046-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:163aa01ed7e6bf51a6d843c49df388545ba06c1a60448f061a160ebac3ea4459
3
- size 173901095
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0277c394a46e19d8386fdf7b5223f4c6300682fa496e6d01e4e5d3abc942f850
3
+ size 173873252
data/train-00050-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:6d13638e72fcfacf697f3f59d2abed4b402c304e69ca70012608a854ddf0bc4a
3
- size 175251771
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f617300d9dddf2af64fc606c6be9c9b8182c66a898a2a6a0633156031edbdfdc
3
+ size 175249899
data/train-00052-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:6729178fcc6015f43428f0be6a3295e8cc5932213a272a74cd1dfd7d864ff378
3
- size 172649245
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:767b7413c3394b94f1d83c13c8afd0c76dabfa5934c848b8a4b9139d49b71c73
3
+ size 172626929
data/train-00056-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:80d1680d584ccdf6adfb9613a424e341624a86dfea1ad768b9b0d17f1230d735
3
- size 176190899
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8b792cd9fd09e4aec9bcd1430edc0cbd88e63b273d2f2eba1898427d5d1d67aa
3
+ size 176174843
data/train-00060-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:17d22b12fe7dffe47e2deaa43baa53e867adbb44213ca0671d59afaf941e54ff
3
- size 176208975
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1dcfc996d5d329531d8f5bbbc39a853065f96b2cbab1d4531b941309631ab30a
3
+ size 176184977
data/train-00064-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0eec69f6631fee83d2e4fe10f874de9b54a327092e2aa9edfccb1f95b007e4c7
3
- size 176453908
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d924f01bc705be5704c8c69149c2e09b3f34918cd467e1e130d8e34064bfec2f
3
+ size 176439090
data/train-00065-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0df5364b56487775e612bf42e70651a52fcc1faedaa9d2cc1afbaaf11bdf2f2f
3
- size 175651985
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da08b73549c9936ea38cfe33d11399641681484663e02a71ab0c9df258a03cdd
3
+ size 175642773
data/train-00066-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a2d94afe2d1cc4c496005c937efc731530bc2f49935c010cb33693c2f5492472
3
- size 176225010
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a07e075dbd96113ccd633ae974bdaa4c9a036e0e343fee300fa94be30122efc7
3
+ size 176213236
data/train-00068-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c8e2ea734ed1d4b562961c387dd44ee6a53c8fb40dcfd7a222004c2597a800ec
3
- size 174553364
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9956f82e77b28aff5e473922a0abfc5f35cd5822996da48e1716837ade189d73
3
+ size 174539684
data/train-00069-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8787f94007b285b468fa276b49d539d72ac0975ad209623298ae87bfeb0ceaa2
3
- size 174621931
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:03184c115007b58125d4e31f309606873b0751468283dc175050f5393f592bcf
3
+ size 174598146
data/train-00072-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c3bec40980492cf6608454ceb17c0e08720025c3a9148c4debf7d8b02786e605
3
- size 174152835
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d2bfed98c64a6a0c0abb26ced8d570261a5582a2f6bf51a3395f00f52c9f67d9
3
+ size 174130344
data/train-00074-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e35410cddbee410d28c1c9317cdb49695a33def40fa8d868ed5aa9bd0c6b94a5
3
- size 173909303
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9db7b54cf6fe0d09051980f7a1e5d460c07ce1a9ac41015263f22ae5ad61aab
3
+ size 173898475
data/train-00076-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8070e8baf174771c158492ab7e1c1c3ed6ec5208fd6267f70ec28dc151840fe5
3
- size 174196433
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:235efdd474d596a7b46829e907e802ee84ec3b24bbcb7aacc8cbe6e5e94428be
3
+ size 174183273
data/train-00077-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:63fddf1809bac1ff5c801d3cef0967a0418f8897bf7d154f1fd82dc0f4f0ef7f
3
- size 177286007
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:de7ec87d25ff70f472e31d95c0f54c9a38c2e73f0eec50725ce0fd02d23a5c46
3
+ size 177267802
data/train-00079-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3571aea24915e43fe5603b6838f2789f96868e62ea84247389cf523ad96aaf90
3
- size 173907445
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:35ee67fe898d5af6d6b40826d2348d4e48e2babff6bb21545e79b8a9fa36f27c
3
+ size 173888479
data/train-00080-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3b1654c52c57ba312401bfb16ba681b00c15e4447f2577636f852301546b8553
3
- size 173602521
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ed6278d02fd75241a8f4394eddb30fddf3a10fa2dd42d03625a4a2ab1fda752
3
+ size 173571750
data/train-00089-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:aa47bcf38cb8edb2b4bc1f06a663f1ab3fc8af6b9b0e1ee54e33569e3da30f8b
3
- size 173653124
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e968e742c0c7deb98227dca015dde01f9e98a2d909f826a0dd76fe694fe66d1
3
+ size 173647372
data/train-00096-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0adced99829ee3d28f20c3ebb288cc6cab1a75c047e52ad90224c79bef0a6457
3
- size 173723077
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:56a47c56dacb2dfbfa25ea8fd800f5d106ecaf3fc4d557316599c42d94c6cf0f
3
+ size 173701719
data/train-00102-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d7005c6296f746cefc5b17bd854e0695ed61844d12e5ab7d42a7436122b31de1
3
- size 176138091
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:29ff09d335ab1808397f965483e5a250a9587a42b13462bc2b139476044e7139
3
+ size 176116312
data/train-00103-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:612b94520b29bfd7db4131213f3b26bee4e3c656763595aca912c32617885095
3
- size 176034784
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bcbe629e561a41511264cba7ce866f31dd27a40d885177a65b64b2bae2af5875
3
+ size 176022848
data/train-00106-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3b344e2575bf591345e39dde4c27a6e81b3b1c3dbeaf9c7a9d3149bdc4b3f393
3
- size 173870416
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:75e0e22a0f1edc758476780985d0ef299c8b6db160fc227a9d3c37cab6be1002
3
+ size 173851431
data/train-00108-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e97fb9be31426aa45987237737add975462afdc7f72a1d878f6c790ea40426f3
3
- size 173682306
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3fd9e25c0a1f74305d86c199c268e3ee007c23f7ea453b02226c6b3312a0dd75
3
+ size 173665522
data/train-00109-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f79edc3fce8ec75722acb300546d5fcd6012e956a1e53e2f1ae23b8e517fe4e6
3
- size 174503185
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd36bc1516a14711cfa6d4852b558c1443c4109064842acd4504b17b338af45d
3
+ size 174473702
data/train-00111-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d96a01d61062a3614250fc9e8e7153593d7a92a277abff07368f642376c49c4a
3
- size 174037465
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:043aaab74e29ca6067587d5bf1e39bf6cc84fc128b08976bb3ce72d9950a7a0c
3
+ size 174023569
data/train-00112-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c00113bb691dd14ed3eca6310ec6043cbf5b385e8d3643c666cabcf12102e47c
3
- size 174368012
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1db6ff3e49e47e39cf32d6b6dbc3b53a0114af7c4d3206d0a4613002b61b3fc8
3
+ size 174352713
data/train-00113-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:390de47fa7c12c39b70f7a93a74bdd7b2ab96fc514b1a0208c43d745c673c8ba
3
- size 174262906
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:505c57e466f32e01ffb8dacf92c876f1c6afab0ff8ceb71b8016b11e545d5d64
3
+ size 174248299
data/train-00116-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:adacbcd57a1d76e76a884010b0c4f8825b0d6d886f5271b517b3b855bcdbfbab
3
- size 175018312
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9b75bd154efad3e22c58779100582ac4deba129ba1ac5d1208ffba242740119
3
+ size 174989187
data/train-00120-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:71f6070fb2558ff7ac0f77210fcad6c4aab7c236251609cd1b32d4a4adb23821
3
- size 177713048
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:616b05d5f322508c7c3d6e9a28d3da8d765a3f17fc525537907953506e6e5179
3
+ size 177704464
data/train-00121-of-00128.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d6e38ff707021981673afae7f0580fc1749bf1d93453f2c9f1bc8a20c9616932
3
- size 173451337
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1304dbd571f5d09f808d9b5c588b22c855ad857906491cb21baf331e7ab754fd
3
+ size 173446793
meta/drop_identity_residual.py ADDED
@@ -0,0 +1,230 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Remove 63 residual identity-leak rows from allenai/Dolci-Think-SFT-Clean-v2.
2
+
3
+ The rows (2026-11-07):
4
+ - 61 Aya rows whose reasoning has the model call itself DeepSeek R1 in Chinese or Japanese
5
+ (such as "作为DeepSeek R1,我应该…" or "私はDeepSeek R1として…"); one of them (_74500) also says it
6
+ in the answer. These are DeepSeek-R1 generations.
7
+ - 2 Dolci Think Precise IF rows (QwQ-32B generations) in which the model says it was built by
8
+ Alibaba Group (_26826, answer) or trained on Alibaba Group data (_834, reasoning).
9
+ Clean-v1's identity filter (A_identity_leak) missed them. They were found by scanning every
10
+ assistant turn of Clean-v2 for any mention of DeepSeek or Qwen, and for QwQ, Tongyi, 通义, 千问,
11
+ 阿里云, 阿里巴巴, "Alibaba Group" and "built by Alibaba"-style phrases, then reading each of the 122
12
+ matching rows by hand. The other 59 are kept. They use the names as a company or topic (Jack Ma,
13
+ Alipay), in R1's own hashtags or promo codes, or as letters inside other words ("qwq" in a puzzle,
14
+ Zulu "emgwaqweni"). A second pass read the 174 further rows that mention plain "Alibaba" or
15
+ contain "qwq" inside a word, and the 41 rows matching other spellings (深度求索, ディープシーク,
16
+ アリババ, ...) or self-claims naming ChatGPT, GPT-4, Claude, Gemini or Llama; none is a
17
+ self-identification.
18
+
19
+ Shards holding none of the rows are left byte-identical. Shards holding some are rewritten with the
20
+ rows removed, in the same row order, with zstd and 2,000-row row groups like the original build.
21
+ meta/drop_ids.parquet gains the 63 ids under reason R_identity_leak_residual, so the source
22
+ (allenai/Dolci-Think-SFT @ 7668c638) minus drop_ids.parquet still reproduces the dataset exactly.
23
+
24
+ Token counts use the build's tokenizer and chat template (allenai/dolma2-tokenizer-olmo35 @
25
+ 8b971706), encoding with its tokenizer.json through `tokenizers`. Under transformers >= 5,
26
+ AutoTokenizer loads this tokenizer-only repo with a plain ByteLevel pre-tokenizer and gives
27
+ different counts (9 to 184 more on these 63 rows); the tokenizer.json encoder reproduced the
28
+ published `tokens` exactly on all 1,745 identity and cutoff rows Clean-v1 dropped, and on a
29
+ separate check of 2,102 dropped rows from eight source row groups.
30
+
31
+ Usage: python -I drop_identity_residual.py DATA_DIR OLD_META_DIR OUT_DIR
32
+ DATA_DIR local copy of data/*.parquet at BASE_REV (checked against the hub's LFS sha256)
33
+ OLD_META_DIR meta/ at BASE_REV (drop_ids.parquet, manifest.json, kimi_k3_contradictory_ids.parquet)
34
+ OUT_DIR receives data/<rewritten shards> and meta/{drop_ids.parquet,manifest.json}
35
+ """
36
+
37
+ import hashlib
38
+ import json
39
+ import sys
40
+ from collections import Counter
41
+ from pathlib import Path
42
+
43
+ import pyarrow as pa
44
+ import pyarrow.compute as pc
45
+ import pyarrow.parquet as pq
46
+ from huggingface_hub import HfApi, hf_hub_download
47
+ from tokenizers import Tokenizer
48
+ from transformers import AutoTokenizer
49
+
50
+ REPO = "allenai/Dolci-Think-SFT-Clean-v2"
51
+ BASE_REV = "b4eaad204084de5ec0c1df41e41f0122dd4a9c70" # data identical to ec947927 / b7eac239
52
+ BASE_ROWS = 2_050_381
53
+ N_SHARDS = 128
54
+ TOK, TOK_REV = "allenai/dolma2-tokenizer-olmo35", "8b9717061fae09d5be814373d189919a62a9a00d"
55
+ REASON = "R_identity_leak_residual"
56
+
57
+ IDS = [
58
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_10180",
59
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_11798",
60
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_12654",
61
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_14227",
62
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_15705",
63
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_18664",
64
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_23057",
65
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_24887",
66
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_25861",
67
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_27241",
68
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_28318",
69
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_28440",
70
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_28789",
71
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_31985",
72
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_34997",
73
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_43722",
74
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_44700",
75
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_44962",
76
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_45464",
77
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_4605",
78
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_46723",
79
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_46831",
80
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_50264",
81
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_53383",
82
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_5650",
83
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_57151",
84
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_57718",
85
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_60896",
86
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_62873",
87
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_64458",
88
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_645",
89
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_64653",
90
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_65578",
91
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_66367",
92
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_66431",
93
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_69088",
94
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_69414",
95
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_69539",
96
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_7220",
97
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_74500",
98
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_75349",
99
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_7704",
100
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_77608",
101
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_79291",
102
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_80702",
103
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_82712",
104
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_82850",
105
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_8319",
106
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_85213",
107
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_85577",
108
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_87600",
109
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_87848",
110
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_87970",
111
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_88802",
112
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_8963",
113
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_89674",
114
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_92770",
115
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_93870",
116
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_95987",
117
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_96068",
118
+ "aya-100k-r1-format-filtered-keyword-filtered-filter-datecutoff-ngram-filtered_96428",
119
+ "if_qwq_reasoning_verified_filtered_decontam_v2_26826",
120
+ "if_qwq_reasoning_verified_filtered_decontam_v2_834",
121
+ ]
122
+
123
+
124
+ def sha256(p):
125
+ h = hashlib.sha256()
126
+ with open(p, "rb") as f:
127
+ for b in iter(lambda: f.read(1 << 24), b""):
128
+ h.update(b)
129
+ return h.hexdigest()
130
+
131
+
132
+ def main():
133
+ data_dir, old_meta, out = (Path(a) for a in sys.argv[1:4])
134
+ (out / "data").mkdir(parents=True, exist_ok=True)
135
+ (out / "meta").mkdir(parents=True, exist_ok=True)
136
+ drop = set(IDS)
137
+ assert len(drop) == len(IDS) == 63
138
+
139
+ # ---- inputs are exactly the hub's files at BASE_REV
140
+ info = HfApi().dataset_info(REPO, revision=BASE_REV, files_metadata=True)
141
+ lfs = {s.rfilename: s.lfs.sha256 for s in info.siblings if s.rfilename.startswith("data/")}
142
+ files = sorted(data_dir.glob("*.parquet"))
143
+ assert len(files) == len(lfs) == N_SHARDS
144
+ for f in files:
145
+ assert sha256(f) == lfs[f"data/{f.name}"], f"{f.name} differs from {REPO}@{BASE_REV}"
146
+ man = json.loads((old_meta / "manifest.json").read_text())
147
+ assert sha256(old_meta / "drop_ids.parquet") == man["files"]["drop_ids.parquet"]
148
+ assert sha256(old_meta / "kimi_k3_contradictory_ids.parquet") == man["files"]["kimi_k3_contradictory_ids.parquet"]
149
+ old_drops = pq.read_table(old_meta / "drop_ids.parquet")
150
+ old_list = old_drops.column("id").to_pylist()
151
+ assert old_list == sorted(old_list), "old drop list not sorted by id"
152
+ assert old_drops.num_rows == man["rows_dropped"] and not drop & set(old_list)
153
+
154
+ # ---- filter shards in place
155
+ tok = AutoTokenizer.from_pretrained(TOK, revision=TOK_REV) # chat template only
156
+ enc = Tokenizer.from_file(hf_hub_download(TOK, "tokenizer.json", revision=TOK_REV))
157
+ base_ids, new_ids, rewritten, removed = [], [], [], {}
158
+ for f in files:
159
+ t = pq.read_table(f)
160
+ ids = t.column("id").to_pylist()
161
+ base_ids += ids
162
+ hit = pc.is_in(t.column("id"), value_set=pa.array(IDS))
163
+ if not pc.any(hit).as_py():
164
+ new_ids += ids
165
+ continue
166
+ for r in t.filter(hit).to_pylist():
167
+ text = tok.apply_chat_template(r["messages"], tokenize=False)
168
+ removed[r["id"]] = (r["source"], len(enc.encode(text, add_special_tokens=False).ids))
169
+ kept = t.filter(pc.invert(hit))
170
+ pq.write_table(kept, out / "data" / f.name, compression="zstd", row_group_size=2000)
171
+ back = pq.read_table(out / "data" / f.name)
172
+ assert back.schema.equals(t.schema) and back.column("id").to_pylist() == [i for i in ids if i not in drop]
173
+ new_ids += back.column("id").to_pylist()
174
+ rewritten.append(f.name)
175
+ assert len(base_ids) == BASE_ROWS == len(set(base_ids)) and drop <= set(base_ids)
176
+ assert set(removed) == drop
177
+ assert new_ids == [i for i in base_ids if i not in drop] and len(new_ids) == BASE_ROWS - 63
178
+
179
+ # ---- drop list: old rows unchanged, 63 appended, sorted by id like the original build
180
+ add = pa.table(
181
+ {
182
+ "id": sorted(drop),
183
+ "reasons": [REASON] * 63,
184
+ "source": [removed[i][0] for i in sorted(drop)],
185
+ "tokens": pa.array([removed[i][1] for i in sorted(drop)], pa.int64()),
186
+ "aligned_full_v1": pa.array([None] * 63, pa.int64()),
187
+ },
188
+ schema=old_drops.schema,
189
+ )
190
+ new_drops = pa.concat_tables([old_drops, add])
191
+ new_drops = new_drops.take(pc.sort_indices(new_drops, sort_keys=[("id", "ascending")]))
192
+ pq.write_table(new_drops, out / "meta" / "drop_ids.parquet", compression="zstd")
193
+
194
+ # ---- manifest
195
+ t63 = sum(n for _, n in removed.values())
196
+ src_rows, src_tok = Counter(), Counter()
197
+ for s, n in removed.values():
198
+ src_rows[s] += 1
199
+ src_tok[s] += n
200
+ by_src = {d["source"]: d for d in man["dropped_by_source"]}
201
+ for s in src_rows:
202
+ by_src[s]["rows"] += src_rows[s]
203
+ by_src[s]["tokens"] += src_tok[s]
204
+ man["rows_dropped"] += 63
205
+ man["rows_kept"] -= 63
206
+ man["tokens_dropped"] += t63
207
+ man["tokens_kept"] -= t63
208
+ man["tokens_dropped_identity_residual"] = t63
209
+ man["filters"][REASON] = 63
210
+ man["filters"] = dict(sorted(man["filters"].items()))
211
+ man["dropped_by_source"] = [by_src[s] for s in sorted(by_src)]
212
+ man["files"]["drop_ids.parquet"] = sha256(out / "meta" / "drop_ids.parquet")
213
+ man["identity_residual"] = {
214
+ "date": "2026-11-07",
215
+ "base_revision": BASE_REV,
216
+ "base_rows": BASE_ROWS,
217
+ "rows": 63,
218
+ "tokens": t63,
219
+ "by_source": {s: {"rows": src_rows[s], "tokens": src_tok[s]} for s in sorted(src_rows)},
220
+ "rewritten_shards": rewritten,
221
+ "token_encoder": f"{TOK}@{TOK_REV} chat template, encoded with its tokenizer.json via `tokenizers`",
222
+ "script": "meta/drop_identity_residual.py",
223
+ }
224
+ assert man["rows_kept"] == len(new_ids) and man["rows_dropped"] == new_drops.num_rows
225
+ (out / "meta" / "manifest.json").write_text(json.dumps(man, indent=1, ensure_ascii=False))
226
+ print(json.dumps({k: man[k] for k in ["rows_dropped", "rows_kept", "tokens_dropped", "tokens_kept", "identity_residual"]}, indent=1, ensure_ascii=False))
227
+
228
+
229
+ if __name__ == "__main__":
230
+ main()
meta/drop_ids.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:24708acc1e54bc2e208027775fd00066eb369893e69e722e1a6682c9305e7726
3
- size 1274341
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0bf79e5a153275026fad5e4182e9e4a1e165151d34996496570a10564e002e84
3
+ size 1275176