HuggingMicah/deepmath-103k-10k-sft-nothink

DeepMath-103K · 10K Non-Thinking SFT Subset

A 10,138-example stratified subset of zwhe99/DeepMath-103K, reformatted as ChatML for supervised fine-tuning of mathematical reasoning.

Each target is a single chain-of-thought ending in \boxed{answer}, with no <think> tags: the sample is meant to be rendered with enable_thinking=False, where the Qwen chat template supplies the empty thinking block itself.

Files

File	Rows
train.jsonl	9,634
val.jsonl	504

One record:

{
  "id": "dm002083",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "... Therefore, the answer should be \\boxed{3}."}
  ],
  "meta": {
    "question": "...",
    "final_answer": "3",
    "difficulty": 6.5,
    "topic": "Mathematics -> Algebra -> Algebra -> Polynomial Operations",
    "topic_bucket": "Algebra",
    "source": "zwhe99/DeepMath-103K",
    "solution_field": "r1_solution_1",
    "target_format": "r1_cot_pre_think"
  }
}

meta.question duplicates the user turn for convenience; drop meta if your collator does not ignore unused columns.

How it was built

Starting from all 103,022 rows:

1. Difficulty window 4.5 ≤ d < 8.5 → 87,765 rows.
2. Answer consistency: the last \boxed{} in r1_solution_1 must match final_answer (numeric tolerance, fractions and percentages normalized). Pass rate over the whole corpus: 98.97%.
3. Not truncated: the solution must end with a boxed answer.
4. Length 3,000–50,000 characters.
5. Deduplicate by normalized question (punctuation, whitespace, case stripped) → 82,749 rows.
6. Stratified sampling over (topic, difficulty) cells with a 50% per-cell cap, seed 42.
7. Split by question, 95/5, so no question appears in both splits.
8. Target normalization: DeepSeek-R1 solutions carry a </think> boundary — a long CoT ending in \boxed{}, followed by a second, condensed restatement of the same solution. Everything from </think> onward was removed, keeping the long CoT. 45 rows where the model declared the problem statement incomplete were dropped, plus 25 rows whose target did not end in \boxed{}.

Difficulty quota was 20 / 30 / 30 / 20 over the four bins; realized exactly 20.0 / 30.0 / 30.0 / 20.0.

Topic quota was adjusted to what the corpus can actually supply:

Bucket	Target	Realized	Available in window	Sampling rate
Algebra	24%	24.5%	26,696	9.4%
Calculus	14%	14.7%	26,577	5.6%
Number Theory	14%	13.8%	4,940	28.6%
Geometry	14%	12.9%	6,042	21.8%
Stats	12%	12.0%	5,831	21.0%
Precalculus	8%	8.1%	8,285	9.9%
Discrete	8%	8.0%	4,817	17.0%
Other	4%	4.0%	3,158	13.1%
Differential Equations	2%	2.0%	1,419	14.4%

Number Theory was originally planned at 20%, but the corpus only has 4,940 rows in the window — 20% would have consumed 41% of the whole topic.

Realized distribution (train)

Bucket	Rows
Algebra	2,360
Calculus	1,412
Number Theory	1,331
Geometry	1,249
Stats	1,151
Precalculus	777
Discrete Mathematics	772
Other	388
Differential Equations	194

Difficulty: 4.5 → 415, 5.0 → 1,509, 5.5 → 1,325, 6.0 → 1,565, 6.5 → 1,423, 7.0 → 1,465, 7.5 → 689, 8.0 → 1,243.

Target length (characters): p5 = 4,574, median = 12,021, p95 = 35,287, max = 48,375. About 43M tokens per epoch at ~3.5 characters per token.

Usage

from datasets import load_dataset
from transformers import AutoTokenizer

ds = load_dataset("HuggingMicah/deepmath-103k-10k-sft-nothink")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

text = tok.apply_chat_template(
    ds["train"][0]["messages"],
    tokenize=False,
    enable_thinking=False,          # empty <think> block comes from the template
)

The trailing \boxed{} makes answer extraction a simple regex over the generated text.

Caveats

• This is not ROSS. It is a plain full-sequence SFT set distilled from external R1 solutions, equivalent to a positive-rollout / rejection-sampling baseline. It contains no model-generated rollouts and no token-level supervision mask.
• Roughly 15% of questions are phrased as proofs ("prove that ..."), but all of them were retained only when the solution carried a verifiable boxed answer. Expect the model to pick up some proof-style phrasing.
• Targets are long relative to Qwen's usual no-thinking outputs. With a 16K output budget, a small fraction of generations may hit the cap.
• Only r1_solution_1 is used. The other two solutions are not included, to avoid selection bias in this first version.

Attribution

• Base corpus: zwhe99/DeepMath-103K (MIT), which is decontaminated against common benchmarks.
• Solutions: generated by DeepSeek-R1, as distributed in DeepMath-103K.
• Filtering, normalization, and sampling: this repository.

License

MIT, inherited from DeepMath-103K.