Methods › General › Distillation › SFT
Shrink and Fine-Tune
SFT
Introduced by Sam Shleifer et al. in Pre-trained Summarization Distillation
archive 2025-07-28 Description, source and code snippet are the archive's method entry.
Shrink and Fine-Tune, or SFT, is a type of distillation that avoids explicit distillation by copying parameters to a student student model and then fine-tuning. Specifically it extracts a student model from the maximally spaced layers of a fine-tuned teacher. Each layer l ∈L′ is copied fully from L. For example, when creating a BART student with 3 decoder layers from the 12 encoder layer 12 decoder layer teacher, we copy the teacher’s full Encᴸ and decoder layers 0, 6, and 11 to the student. When deciding which layers to copy, we break ties arbitrarily; copying layers 0, 5, and 11 might work just as well. When copy only 1 decoder layer, we copy layer 0. This was found this to work better than copying layer 11. The impact of initialization on performance is measured experimentally in Section 6.1. After initialization, the student model continues to fine-tune on the summarization dataset, with the objective of minimizing ℒ_(Data).
Papers archive 2025-07-28
30 shown of 415, newest first. Repository counts are the archive's code-links table. A Syntology line states what Syntology ran from that paper's harvested code; it is per sample and not a correctness claim.
-
A Practical Two-Stage Recipe for Mathematical LLMs: Maximizing Accuracy with SFT and Efficiency with Reinforcement Learning 11 Jul 2025 · 1 repository · arXiv:2507.08267Syntology ran 0 of 6 samples · 6 unverified · 6 pointer-only (licence)
-
InvestAlign: Overcoming Data Scarcity in Aligning Large Language Models with Investor Decision-Making Processes under Herd Behavior 9 Jul 2025 · 1 repository · arXiv:2507.06528
-
AutoTriton: Automatic Triton Programming with Reinforcement Learning in LLMs 8 Jul 2025 · 1 repository · arXiv:2507.05687
-
CogniSQL-R1-Zero: Lightweight Reinforced Reasoning for Efficient SQL Generation 8 Jul 2025 · 0 repositories · arXiv:2507.06013
-
Reinforcement Fine-Tuning Naturally Mitigates Forgetting in Continual Post-Training 7 Jul 2025 · 0 repositories · arXiv:2507.05386Syntology ran 3 of 10 samples · 7 unverified · 10 pointer-only (licence)
-
SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning 30 Jun 2025 · 1 repository · arXiv:2506.24119Syntology ran 2 of 9 samples · 7 unverified
-
Complexity-aware fine-tuning 26 Jun 2025 · 1 repository · arXiv:2506.21220
-
LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement Learning 23 Jun 2025 · 1 repository · arXiv:2506.18841Syntology ran 6 of 14 samples · 8 unverified
-
Smart-LLaMA-DPO: Reinforced Large Language Model for Explainable Smart Contract Vulnerability Detection 23 Jun 2025 · 0 repositories · arXiv:2506.18245
-
RLRC: Reinforcement Learning-based Recovery for Compressed Vision-Language-Action Models 21 Jun 2025 · 0 repositories · arXiv:2506.17639
-
Differentiation-Based Extraction of Proprietary Data from Fine-Tuned LLMs 20 Jun 2025 · 0 repositories · arXiv:2506.17353
-
Refine-POI: Reinforcement Fine-Tuned Large Language Models for Next Point-of-Interest Recommendation 19 Jun 2025 · 0 repositories · arXiv:2506.21599
-
Massive Supervised Fine-tuning Experiments Reveal How Data, Layer, and Training Factors Shape LLM Alignment Quality 17 Jun 2025 · 0 repositories · arXiv:2506.14681
-
AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy 16 Jun 2025 · 0 repositories · arXiv:2506.13284
-
Ego-R1: Chain-of-Tool-Thought for Ultra-Long Egocentric Video Reasoning 16 Jun 2025 · 0 repositories · arXiv:2506.13654
-
Metis-RISE: RL Incentivizes and SFT Enhances Multimodal Reasoning Model Learning 16 Jun 2025 · 1 repository · arXiv:2506.13056
-
QFFT, Question-Free Fine-Tuning for Adaptive Reasoning 15 Jun 2025 · 1 repository · arXiv:2506.12860Syntology ran 3 of 14 samples · 11 unverified
-
VGR: Visual Grounded Reasoning 13 Jun 2025 · 0 repositories · arXiv:2506.11991
-
Vision Matters: Simple Visual Perturbations Can Boost Multimodal Math Reasoning 11 Jun 2025 · 1 repository · arXiv:2506.09736
-
Eliciting Fine-Tuned Transformer Capabilities via Inference-Time Techniques 9 Jun 2025 · 1 repository · arXiv:2506.08060
-
Learning What Reinforcement Learning Can't: Interleaved Online Fine-Tuning for Hardest Questions 9 Jun 2025 · 1 repository · arXiv:2506.07527
-
Nearness of Neighbors Attention for Regression in Supervised Finetuning 9 Jun 2025 · 1 repository · arXiv:2506.08139
-
Learning to Clarify by Reinforcement Learning Through Reward-Weighted Fine-Tuning 8 Jun 2025 · 0 repositories · arXiv:2506.06964
-
SDE-SQL: Enhancing Text-to-SQL Generation in Large Language Models via Self-Driven Exploration with SQL Probes 8 Jun 2025 · 0 repositories · arXiv:2506.07245
-
Writing-RL: Advancing Long-form Writing via Adaptive Curriculum Reinforcement Learning 6 Jun 2025 · 1 repository · arXiv:2506.05760
-
MINT-CoT: Enabling Interleaved Visual Tokens in Mathematical Chain-of-Thought Reasoning 5 Jun 2025 · 1 repository · arXiv:2506.05331
-
On the Mechanism of Reasoning Pattern Selection in Reinforcement Learning for Language Models 5 Jun 2025 · 0 repositories · arXiv:2506.04695
-
Crowd-SFT: Crowdsourcing for LLM Alignment 4 Jun 2025 · 0 repositories · arXiv:2506.04063
-
RoboRefer: Towards Spatial Referring with Reasoning in Vision-Language Models for Robotics 4 Jun 2025 · 0 repositories · arXiv:2506.04308
-
Corrigibility as a Singular Target: A Vision for Inherently Reliable Foundation Models 3 Jun 2025 · 0 repositories · arXiv:2506.03056
Tasks archive 2025-07-28
20 shown of 206 tasks the archive attaches to papers tagged with this method, by distinct papers. A task without a page in the catalog is plain text.
Usage over time archive 2025-07-28
Components: the archive holds no method-to-method composition, so PwC's Components table cannot be rebuilt; the Papers list carries no Results column for the same reason (the archive does not join its leaderboard rows to method tags).
Categories archive 2025-07-28
Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections