Dolci-Think-SFT-Clean-v1

A public filtered research snapshot of the 32B post-training dataset allenai/Dolci-Think-SFT, built from revision 7668c638cc84100951973b456069a1a462d6d915 during an October 2026 audit. It removes 63,614 of the source's 2,253,684 rows. Those rows carry 831,168,603 tokens, 3.67% of the source's tokens.

• Size (default): 2,190,070 rows and 21,831,528,207 tokens, using the historical build's olmo35 tokenizer and template.
• Schema and order: same as the source (messages, id, source), in source row order. Filtering removes whole rows; retained message content is unchanged.
• Scope: Clean is the snapshot's historical name, not a guarantee of correctness, safety, or complete benchmark decontamination.

Load

from datasets import load_dataset

data = load_dataset(
    "allenai/Dolci-Think-SFT-Clean-v1",
    "default",
    revision="d03e7c65f87149c39945da78714a7e0e2e13f173",
    split="train",
    streaming=True,
)
example = next(iter(data))

This example pins the original data revision for reproducible work.

Removed rows

The table records the original audit's filter labels and judgments. A row can carry several reasons. The 63,614 rows are unique; the counts below sum to more.

Reason	Rows	What it is
A_tierA_contam	301	Benchmark contamination identified by the audit's Tier-A filter
A_think_empty	3,264	Empty reasoning
A_identity_leak	666	Identity leaks
A_malformed	270	No </think>, not exactly one <think>, or an empty answer
C_ot3math_truncated	30,038	OT3-Math answers cut off
B_degenerate_tail_C4	239	Repetition-collapse tails
B_ctrl_token_and_stop_strings	789	Literal control tokens or stop strings in content
B_ot3_code_stem_capped	11,633	OT3 Code/STEM answers truncated at the teacher's length cap. Every prompt keeps a complete sibling
B_if_harmful_complied	3,470	Instruction-following rows that comply with harmful WildGuardMix/WildJailbreak seeds
B_tablegpt_wrong	630	Wrong TableGPT answers
B_dated_cutoff_noidentity	1,021	Conflicting dated knowledge-cutoff claims
B_olmo_identity_rows	58	Stale Olmo identity rows
B_chat_cap2	11,754	Repeated chat prompts beyond 2 per key

Known limitations

This snapshot is not fully benchmark-decontaminated. The 301-row historical contamination filter does not cover all subsequently reported overlaps. See the source dataset's public benchmark-contamination discussion for findings involving MATH, multi-turn IFEval (ifeval_mt), IFBench, MMLU, and GPQA. The GPQA finding is a 13-gram overlap signal, not confirmed exact contamination. That discussion describes the source dataset; its counts are not a measurement of this filtered subset. Check overlap against the benchmarks used in your own evaluation.

The original audit also identified unresolved issues that this row filter does not repair:

• [CUTOFF_DATE] placeholders.
• Hidden safety-system-prompt issues in WildGuardMix, WildJailbreak, and CoCoNot rows.
• Wrong-language Aya rows.

The filters reflect an audit's judgments and do not establish that every retained answer is correct or every removed answer is unusable. The optional ill-posedness selector described below was not applied to any config.

Configs

• default: the filtered corpus.
• arm-T-ot3cap4: historical experimental treatment subset: 971,930 rows, 7,648,640,122 tokens.
  • OpenThoughts3 (OT3) Math/STEM/Code rows are capped at 4 responses per (source, prompt), with a seeded random pick. Every OT3 prompt is kept.
  • Non-OT3 rows are sampled at a rate of 12,339,642,368 / 21,831,528,207 = approximately 0.565. The numerator is the historical experiment's token budget.
• arm-R-random-matched: historical experimental control subset: 767,320 rows, 7,648,635,935 tokens.
  • Random rows of the filtered corpus, token-matched to T.

T and R were constructed for an OT3 response-repetition experiment, with an intended training schedule of exactly one pass each. Seed 34521. Their inclusion does not establish a training benefit; they are experimental subsets rather than recommended training recipes.

Provenance and reproducibility

The exact default row selection can be reproduced from the pinned public source revision above by excluding IDs in meta/drop_ids.parquet and preserving source shard and row order. The drop list contains 63,614 unique IDs with reason labels, source names, and historical token counts. Its SHA-256 is 2fdadcb64b51022877049a7fa48732117caac82bcf6cf5d0d377cb7a18255bd0.

The published ID lists likewise specify exact membership of T and R. Regenerating the original filter judgments, token counts, or random selections from scratch requires audit feature and selector artifacts that are not published here. The archived scripts refer to that original build environment and are not standalone reproduction scripts. The row-selection lists remain usable independently of those artifacts.

File in meta/	Contents
drop_ids.parquet	One row per dropped ID: id, reasons, source, aligned_full token count
manifest.json	Source revision, counts, filter counts, per-source removal counts, and SHA-256 hashes of the original local build files; the original corpus-file hash is not a hash of the uploaded shards
build.py, make_arms.py, hf_upload.py	Archived scripts for the original filtering, experimental subsets, and upload
arms_summary.json	Counts and token totals for the experimental subsets
T_ot3cap4_ids.parquet, R_random_matched_ids.parquet	ID-only membership lists for T and R
ILL_ot3math_illposed_ge05_drop_ids.parquet	ID-only optional selector: 461,760 OT3-Math rows whose prompts at least half of the historical judgments called ill-posed. Not applied to any config.

The historical audit estimated the ILL selector's precision at about 78% for answer-changing flaws; the other approximately 22% were benign or degenerate prompts. This is a limitation of the unapplied selector, not a correctness estimate for the corpus.

License

Licensing follows allenai/Dolci-Think-SFT: ODC-BY overall, with per-source licenses listed in that card.