Small-Data Reasoning Fine-Tuning and Distillation
SFT on hundreds to thousands of curated reasoning traces from a stronger model.
Tier Grades
Tier A: Frontier AI labs
feasible
high confidence
Frontier labs hold the strong base models these methods depend on, so the topic is within their reach.
Tier B: Indonesian AI labs
feasible
high confidence
s1 needed 16 H100 for 26 minutes and 1,000 examples, far below the Sahabat-AI training scale b already reached.
Tier C: Institution units such as ITB
feasible
high confidence
Universities and nonprofits produced s1, and 16 H100 for 26 minutes is a few GPU-hours, within the academic 1 to 8 GPU budget given more wall time.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
medium confidence
1,000 curated examples and a few GPU-hours on an open base fit a proof-of-concept budget, though the notes give no direct proof-of-concept anchor.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- Computes1: 16 H100 for 26 minutes on Qwen2.5-32B-Instruct; Absolute Zero compute not reported.
- Datas1: 1,000 curated examples; LIMO: about 1% of prior training data; Absolute Zero: self-play with a code executor, no external data.
- Team size (proxy)Proxy only: s1 author institutions are Stanford, UW, AI2 and Contextual AI; author count not given in the notes.
- Time horizonDays to weeks.
- Shelf-life signalsAll three exemplars depend on an existing strong open base model, so they track frontier open releases (topic 24)
- Pipeline roleConsumes Tier A's open bases and reasoning traces; Tier C and Tier B produce small specialised reasoning models that Tier D turns into proofs of concept.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
Absolute Zeroreasoningexemplar
LIMOreasoningexemplar
s1reasoningexemplar
OpenThinker-7Breasoningproduced by
DeepSeek-R1-Distill-Llama-70Breasoningproduced byDeepSeek-R1-Distill-Llama-8Breasoningproduced byDeepSeek-R1-Distill-Qwen-1.5Breasoningproduced byDeepSeek-R1-Distill-Qwen-14Breasoningproduced byDeepSeek-R1-Distill-Qwen-32Breasoningproduced byDeepSeek-R1-Distill-Qwen-7Breasoningproduced byDeepSeek-R1reasoningused by
Qwen2.5-32B-InstructLLMused byLinks
R&D TopicSmall-Data Reasoning Fine-Tuning and Distillation
ModelsAbsolute ZeroTsinghua University, reasoningLIMOShanghai Jiao Tong University, reasonings1Stanford University, reasoningQwen2.5-32B-InstructAlibaba (Qwen), LLM+8 more
Absolute ZeroTsinghua University, reasoningLIMOShanghai Jiao Tong University, reasonings1Stanford University, reasoningQwen2.5-32B-InstructAlibaba (Qwen), LLM+8 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 12
DeepSeek-R1DeepSeek, reasoningused byDeepSeek-R1-Distill-Llama-8BDeepSeek, reasoningproduced byDeepSeek-R1-Distill-Qwen-1.5BDeepSeek, reasoningproduced byDeepSeek-R1-Distill-Qwen-7BDeepSeek, reasoningproduced bys1Stanford University, University of Washington, Allen AI, Contextual AI, reasoningexemplarDeepSeek-R1-Distill-Llama-70BDeepSeek, reasoningproduced byDeepSeek-R1-Distill-Qwen-14BDeepSeek, reasoningproduced byDeepSeek-R1-Distill-Qwen-32BDeepSeek, reasoningproduced byQwen2.5-32B-InstructAlibaba (Qwen), LLMused byAbsolute ZeroTsinghua University, reasoningexemplarOpenThinker-7BOpen Thoughts, reasoningproduced byLIMOShanghai Jiao Tong University, reasoningexemplarPapers
- LIMA: Less Is More for AlignmentChunting Zhou et al., arXiv preprint, 2023 founding
- STaR: Bootstrapping Reasoning With ReasoningEric Zelikman et al., arXiv preprint, 2022 founding
- s1: Simple test-time scalingNiklas Muennighoff et al., arXiv preprint, 2025 state of the art
- LIMO: Less is More for ReasoningYixin Ye et al., COLM 2025 (per arXiv comment), 2025 state of the art
- Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model SizesCheng-Yu Hsieh et al., ACL 2023 (per arXiv comment), 2023 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- $100K or 100 Days:Trade-offs when Pre-Training with Academic Resources arxiv.org https://arxiv.org/html/2410.23261v1
- s1: Simple test-time scaling arxiv.org https://arxiv.org/html/2501.19393v1
- LIMO: Less is More for Reasoning arxiv.org https://arxiv.org/abs/2502.03387
- Absolute Zero: Reinforced Self-play Reasoning with Zero Data arxiv.org https://arxiv.org/abs/2505.03335
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Small-Data Reasoning Fine-Tuning and Distillation", https://aurion.semai.id/topics/small-data-reasoning-fine-tuning/, records of October 2026.