Releases·Tool Release
1,451 verified max-effort reasoning traces from DeepSeek v4.1 Flash, decontaminated against AIME + MATH-500 — MIT
A community member released a dataset of 1,451 verified, max-effort math reasoning traces generated by DeepSeek V4.1 Flash on Hugging Face under an MIT license. Traces are answer-verified with a sympy grader plus a Qwen3.5-4B judge and decontaminated against MATH-500 and AIME, targeting SFT/distillation of small reasoning models.
0 engagement1 sourceSat, Sep 19, 2026, 11:29 AM
AI Overview
- The dataset ships 11.6M reasoning tokens (~8k avg per trace, max 119k) sourced from AoPS forums and the MATH dataset.
- Traces come from a 2,000-problem stratified sample where DeepSeek V4.1 Flash scored 78.5% pass@1, generating 18.1M tokens in ~9 minutes at concurrency 2500.
- Verification pairs a sympy-based code grader with a locally run Qwen3.5-4B judge that only rescues grader mismatches; wrong or ambiguous traces are excluded.
- Decontamination uses Qwen3-VL-Embedding-8B embeddings, dropping 35 MATH-500 matches and 86 AIME-adjacent traces against a 993-problem 1983–2026 reference.
- Released MIT-licensed on Hugging Face for SFT/distillation of small reasoning models, with 87 traces exceeding 30k thinking tokens usable for long-horizon evals.
- Caveats: answers are verified but reasoning steps are not line-by-line audited, gold answers are inherited from the parent Big-Math-RL-Verified dataset, and all traces share a single teacher.
Entities
Related
Releases·Model Releasefeatured
DeepSeek launches V4.1 Flash: 552B MoE with asymmetric architecture, open weights, V4 Pro phase-out
DeepSeek officially released DeepSeek-V4.1-Flash on September 10, 2026, after an internal API beta opened September 8. The company says the 552B-parameter MoE model uses a new Causal-Encoder-Decoder architecture with asymmetric compute (8B activated on input, 16B on output) and comprehensively outperforms V4 Pro on performance, cost, speed, and time-to-completion, prompting an orderly V4 Pro phase-out. Weights and a tech report are on Hugging Face, and Tencent and OpenCode have integrated it.
DeepSeek officially released DeepSeek-V4.1-Flash on September 10, 2026, after an internal API beta opened September 8. The company says the 552B-parameter MoE model uses a new Causal-Encoder-Decoder architecture with asymmetric compute (8B activated on input, 16B on output) and comprehensively outperforms V4 Pro on performance, cost, speed, and time-to-completion, prompting an orderly V4 Pro phase-out. Weights and a tech report are on Hugging Face, and Tencent and OpenCode have integrated it.
Updated Thu, Sep 10, 2026, 04:05 PM — V4.1-Flash is live on the DeepSeek API, HuggingChat, and OpenCode Go; V4 Pro requests will reroute to Flash from Septemb
36044.4k engagement10 sourcesreddit, hackernews, xThu, Sep 10, 2026, 12:00 AM
Social·Community
Fine-tuning Qwen 3 4B Base on 100 zebra puzzles yielded +31% on MATH-500. 6.5-min (Single H100/H200) reproduction notebook included.
A practitioner fine-tuned Qwen 3 4B Base on just 100 zebra puzzles and reported a +31% gain on MATH-500, sharing the result on Hugging Face. A reproduction notebook is included and runs in about 6.5 minutes on a single H100 or H200, making the experiment easy for others to verify cheaply.
A practitioner fine-tuned Qwen 3 4B Base on just 100 zebra puzzles and reported a +31% gain on MATH-500, sharing the result on Hugging Face. A reproduction notebook is included and runs in about 6.5 minutes on a single H100 or H200, making the experiment easy for others to verify cheaply.
42 engagement1 sourceredditFri, Sep 11, 2026, 02:06 PM
Releases·Tool Release
Hacker shrinks DeepSeek V4 Flash to 57GB, runs it on Mac to write a compiler
A developer packaged DeepSeek V4 Flash 0731 (284B total, 13B active) into a 57GB specialized build that runs on Silicon Macs via a custom engine. The model successfully wrote a minimal C compiler targeting ARM64 in under an hour, demonstrating feasibility of running large MoE models locally.
A developer packaged DeepSeek V4 Flash 0731 (284B total, 13B active) into a 57GB specialized build that runs on Silicon Macs via a custom engine. The model successfully wrote a minimal C compiler targeting ARM64 in under an hour, demonstrating feasibility of running large MoE models locally.
4 engagement2 sourceshackernews, redditSun, Aug 16, 2026, 05:13 PM
Developments·Paper
Magenta: Closing the Loop Between Mathematical Reasoning and Lean Verification
A new arXiv paper introduces Magenta, a training-free agentic pipeline that converts natural-language math problems into Lean 4 statements and machine-checked proofs. It reports 100% accuracy on AIME 2025, AIME 2026, and HMMT February 2026, and solves all six IMO 2026 problems when paired with the K2-Horizon-7B reasoner. The work matters to practitioners because it shows formal verification signals can be integrated into LLM reasoning without retraining.
A new arXiv paper introduces Magenta, a training-free agentic pipeline that converts natural-language math problems into Lean 4 statements and machine-checked proofs. It reports 100% accuracy on AIME 2025, AIME 2026, and HMMT February 2026, and solves all six IMO 2026 problems when paired with the K2-Horizon-7B reasoner. The work matters to practitioners because it shows formal verification signals can be integrated into LLM reasoning without retraining.
0 engagement1 sourcearxivThu, Sep 10, 2026, 09:48 AM