Dolci-Think-SFT-Clean-v2

A filtered snapshot of the 32B post-training dataset allenai/Dolci-Think-SFT, built from revision 7668c638cc84100951973b456069a1a462d6d915. It applies every removal in 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.

• 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).
• Schema and order: same as the source (messages, id, source), in source row order. Filtering removes whole rows. Retained rows are unchanged.
• 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.
• Scope: Clean is the snapshot's 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-v2",
    revision="ec9479277732a7ecf0eb2776ea02e50ddcce8943",
    split="train",
    streaming=True,
)
example = next(iter(data))

This revision is the commit that added the data. Later commits change only documentation.

Removed rows

A row can carry several reasons. The 203,303 rows are unique, so the counts below sum to more.

Reason	Rows	What it is
K_kimi_k3_contradictory	148,210	New in v2. OT3-Math problems a Kimi K3 screen judged self-contradictory or ill-posed (below)
A_tierA_contam	301	Benchmark contamination identified by the v1 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

The A_, B_ and C_ rows are exactly Clean-v1's drop list. See the Clean-v1 card for how they were chosen.

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.

The Kimi K3 screen

Kimi K3 re-solved the OT3-Math prompts. A row is flagged when Kimi's own answer does all of the following:

1. It gives no \boxed{} answer.
2. It ends with language saying the problem has no solution, is contradictory, undefined, underdetermined or ill-posed.
3. It states no final answer.

Rows whose prompt asks for a proof are not flagged. Kimi is used only as a detector. No Kimi text is in this dataset.

On 200 flagged rows drawn uniformly at random and labeled by hand, 187 had a broken problem. That is a precision of 93.5%, 95% Wilson interval [89.2%, 96.2%]. The screen's recall is not measured. Broken math problems that Kimi answered anyway, or that Kimi was not run on, remain.

Training evidence

In a controlled SFT experiment on an Olmo 3.5 base, the Kimi deletion alone was applied to the full source, with the token budget and every other recipe element held fixed. The deletion did not include the Clean-v1 removals. Compared with the undeleted baseline:

• Pooled MATH (7 subjects, n=5,000, T=1.0) rose by 1.69 points, 95% CI [+0.69, +2.69].
• GSM8K rose by 2.20 points.
• PopQA changed by -0.02 points.

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.

Known limitations

Clean-v1's limitations carry over unchanged:

• This snapshot is not fully benchmark-decontaminated. See the source dataset's benchmark-contamination discussion for MATH, multi-turn IFEval, IFBench, MMLU and GPQA findings. Check overlap against the benchmarks you evaluate on.
• [CUTOFF_DATE] placeholders, hidden safety-system-prompt issues in WildGuardMix, WildJailbreak and CoCoNot rows, and wrong-language Aya rows are not repaired.
• Clean-v1's optional ill-posedness selector (ILL_ot3math_illposed_ge05_drop_ids.parquet, 461,760 rows) is still not applied. It and the Kimi screen are different selectors.

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.

Token counts

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.

Clean-v1's card reports aligned_full counts from its audit's build, which are not published. The two counts differ:

• On v1's 63,614 dropped rows, this build's count is 18 tokens higher on 98.4% of rows and differs by other amounts on the rest.
• Totals: 832,330,068 here against 831,168,603 for those rows, and 22,703,597,913 against 22,662,696,810 for the whole source.

So compare v1 and v2 token figures within one scale. The drop_ids.parquet file carries both: tokens (this build) and aligned_full_v1 (v1's, null for rows new in v2).

	Rows	Tokens (this build)
Source	2,253,684	22,703,597,913
Removed, v1 reasons only	55,093	677,213,688
Removed, Kimi only	139,689	2,711,005,640
Removed, both	8,521	155,116,380
Clean-v2	2,050,381	19,160,262,205

Provenance and reproducibility

The exact row selection can be reproduced from the pinned public source revision by excluding the IDs in meta/drop_ids.parquet and keeping source shard and row order.

The build checked the following before uploading:

• Removing Clean-v1's drop list alone (sha256 2fdadcb64b51022877049a7fa48732117caac82bcf6cf5d0d377cb7a18255bd0) reproduces Clean-v1's published default row IDs exactly, in order.
• Removing the Kimi list alone reproduces the dataset the training experiment above used.
• The output keeps the source schema, and its ID sequence is the source's minus the union of the two lists.

File in meta/	Contents
drop_ids.parquet	One row per dropped ID: id, reasons (comma-joined), source, tokens, aligned_full_v1
kimi_k3_contradictory_ids.parquet	The 148,210 Kimi-flagged IDs on their own
manifest.json	Source and v1 revisions, counts, per-filter and per-source removals, token totals and calibration, file hashes
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

Regenerating the Kimi flags from scratch requires the Kimi K3 completions, which are not published here.

License

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