Mephisto-Knowledge_538k

538,861 English knowledge SFT examples generated by Qwen/Qwen3.5-4B in non-thinking (Instruct) mode on the Knowledge prompts of openbmb/UltraData-SFT-2605.

Responses contain no chain-of-thought — thinking was disabled at generation time, so each assistant turn is a direct answer, usually with a short justification.

Companion dataset: Mephisto-IF_172k (instruction-following, same teacher and pipeline).

Read this before training: ref_agrees

Every prompt is multiple-choice, and the source dataset's reference answer states the correct option. That gives a direct correctness signal, and it is published per-row as ref_agrees:

value	rows	meaning
true	269,441	teacher's chosen option matches the reference
false	73,124	teacher contradicts the reference
null	196,296	no letter extractable from one side — unjudged

Measured teacher accuracy is 78.65% on the 342,565 judgeable rows. If you want a high-confidence subset, filter to ref_agrees == true.

These rows were flagged rather than deleted, deliberately: hand-inspection showed false mixes genuine model errors (one misread k_f = 100 as 1000 mid-derivation) with genuinely ambiguous questions where the teacher's reasoning was sound and the reference debatable (a 30-hour half-life "steady state" item where 4 vs 5 half-lives are both defensible). Deleting them would have thrown away the ambiguity along with the errors.

Accuracy falls as answers get longer — the model rambles when unsure:

answer length	n	accuracy
< 100 tokens	69,315	84.2%
100–500 tokens	196,527	79.8%
≥ 500 tokens	76,723	70.8%

So the short, label-only answers are the most reliable rows here, not the low-effort tail. completion_tokens is published per row if you want to slice on it.

Ordering

Rows are shuffled (deterministic, seed 0). The generation driver writes results in completion order, so without shuffling the fastest — i.e. shortest — answers pile up at the head of the file: the first 100 rows were 100% under 50 tokens against a 19.9% corpus rate. Shuffled, the head now matches the corpus, so viewer previews and take(n) on a streaming load are representative.

Format

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "uid": "Knowledge_no_think_0000001",
  "system": "You are a helpful assistant. Answer the user's question accurately, clearly, and concisely.",
  "teacher": "Qwen/Qwen3.5-4B",
  "ref_agrees": true,
  "completion_tokens": 88
}

Answer style: mean 378 completion tokens (p50 290, p90 715). ~31% open with an explicit answer declaration, ~26% explain why the other options are wrong, ~36% contain LaTeX, ~86% use markdown structure.

Generation

Prompts	openbmb/UltraData-SFT-2605, config Knowledge, split no_think
Teacher	Qwen/Qwen3.5-4B, bfloat16, unquantized
Thinking	disabled via chat_template_kwargs: {"enable_thinking": false}
Sampling	temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0 (official Qwen3.5 Instruct-mode recommendation)
Max output	16,384 tokens (--max-model-len 32768)
Hardware	TPU v6e-8, vLLM + tpu-inference, 8 single-chip replicas

This is a partial run. The split has 800,000 prompts; 560,000 were generated before the Spot TPU was preempted. Coverage is a contiguous prefix of the split, and there were 0 request failures across all 560,000.

Filtering

560,000 raw generations → 538,861 kept (96.23%):

Dropped	Rows	Rule
truncated	13,399 (2.39%)	hit the 16,384-token cap
switch echoed	3,055 (0.55%)	model parroted /no_think back
fabricated	2,734 (0.49%)	long answer to a prompt missing its diagram
degenerate	1,733 (0.31%)	zlib compression ratio > 6.0
chinese	218 (0.04%)	>5% CJK characters
kept	538,861

/no_think cleanup. 48.9% of UltraData Knowledge prompts carry a trailing /no_think — Qwen3's soft switch, which Qwen3.5 removed, so it reached the model as meaningless literal text. It is stripped from all published prompts, and the 0.55% of rows where the model echoed it into its answer are dropped.

Degeneracy. Scored with zlib compression ratio rather than n-grams: healthy prose sits at p50 1.99 / p90 2.41, degenerate loops at 10–800x. finish_reason alone does not catch it — the worst offenders repeat a phrase for thousands of tokens and still terminate with stop.

Fabrication. ~0.8% of prompts reference a diagram or structure absent from the text. The model does not refuse; it invents a setup and answers confidently (worst case: 6,936 tokens of IUPAC derivation for a compound it cannot see). Rows that signal missing information and run past 1,000 tokens are dropped. Short answers that simply say "cannot be determined" are kept — that is good behaviour, not a defect.

Known limitations

• The teacher is a 4B model; 21.4% of judgeable answers contradict the reference. Use ref_agrees accordingly.
• 100% multiple-choice. This teaches justified option selection, not open-ended knowledge exposition.
• Partial coverage of the source split (560k of 800k prompts).
• English-only by choice; the source split is bilingual.
• Filtering is heuristic; some low-quality-but-not-degenerate rows remain.

Provenance and license

Prompts from openbmb/UltraData-SFT-2605 (Apache-2.0); responses generated by Qwen/Qwen3.5-4B (Apache-2.0). Released under Apache-2.0.