Card: document the 63 residual identity rows and pin the new data revision
Browse files
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,
|
| 17 |
|
| 18 |
-
- **Size:** 2,050,
|
| 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
|
| 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="
|
| 31 |
split="train",
|
| 32 |
streaming=True,
|
| 33 |
)
|
| 34 |
example = next(iter(data))
|
| 35 |
```
|
| 36 |
|
| 37 |
-
This revision is the commit that
|
| 38 |
|
| 39 |
## Removed rows
|
| 40 |
|
| 41 |
-
A row can carry several reasons. The 203,
|
| 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 |
-
|
|
|
|
|
|
|
|
|
|
|
| 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
|
|
|
|
| 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 |
|