Datasets:

---

allenai

/

Dolci-Think-SFT-Clean-v2

Like 9

Follow

Ai2 7.05k

Written by Claude Code session qwen-identity.

Removes 63 rows in which the model identifies itself as another model, which Clean-v1's identity filter missed: 61 Aya rows where DeepSeek R1 names itself in Chinese or Japanese reasoning ("作为DeepSeek R1，我应该…"), and 2 Dolci Think Precise IF rows (QwQ-32B) saying Alibaba Group built or trained it. Result: 2,050,318 rows, 19,160,089,253 tokens.

Why: the Olmo 3 7B models, trained on data that kept such rows, call themselves Qwen or DeepSeek even under an Olmo system prompt. These rows would teach an Olmo 3.5 SFT the same, under the template's default "You are Olmo" system prompt.

• 47 shards rewritten with the rows removed and order kept; the other 81 are unchanged.
• meta/drop_ids.parquet gains the 63 ids as R_identity_leak_residual, so the source minus drop_ids.parquet still reproduces the data. meta/manifest.json is updated to match.
• meta/drop_identity_residual.py rebuilds this update from b4eaad20 byte for byte.
• The card documents the rows, the scan and two borderline rows that are kept.
• An independent check recomputed the data, the drop list, the token counts and the manifest, and confirmed that all 63 rows are real self-identifications.

The card's Load section pins this PR's data commit. If merging changes commit ids, set revision= to the merge commit.