# lasgroup/SDPO

Reinforcement Learning via Self-Distillation (SDPO)

Repository: https://github.com/lasgroup/SDPO
Canonical: https://ross.abutalabs.com/products/sdpo
Homepage: https://self-distillation.github.io/SDPO
Language: Python
License: Apache-2.0
License Family: permissive
Topics: distillation, llm, reasoning, rl
Last push: 2026-07-01T13:51:47+00:00

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 90, release rhythm 35, longevity 15
- inputs: {"age_days": 221, "days_push": 63, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1075, forks 125 (observed 2026-08-28T04:03:29.188210+00:00)

## What it is
SDPO (Self-Distilled Policy Optimization) is a research library implementing a reinforcement learning framework for post-training large language models. It converts tokenized environment feedback into dense learning signals by distilling the model's own feedback-informed predictions back into the policy, without an external teacher or reward model.

## Use cases
- train llms with reinforcement learning on verifiable rewards
- improve sample efficiency of rl post-training for math and code reasoning
- use rich textual feedback like runtime errors as a dense training signal
- fine-tune a policy model using self-distillation from its own successful rollouts
- accelerate test-time discovery on hard binary-reward tasks

## When to choose
- you are doing research on RLVR or rich-feedback reinforcement learning for LLMs
- you want denser credit assignment than scalar outcome rewards provide
- you need a reference implementation of the SDPO paper

## When to avoid
- you need a production-grade, supported training framework
- you lack GPU resources for LLM post-training
- you want simple supervised fine-tuning without RL

## Facets
- artifact type: library
- maturity: active
- function: llm-training, reinforcement-learning, machine-learning, deep-learning
- domain: large-language-models, machine-learning, reinforcement-learning, artificial-intelligence
- platform: python
- tags: rlvr, self-distillation, policy-optimization, reasoning, research-code, verifiable-rewards, gpu, linux

## Member repositories
- lasgroup/SDPO (main) score 56

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:29.188210+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:53:26.816348+00:00, confidence not recorded.
  - readme: https://github.com/lasgroup/SDPO (fetched 2026-08-28T04:03:29.188210+00:00, sha d29eda2094a2)
  - homepage: https://self-distillation.github.io/SDPO (fetched 2026-08-29T12:55:18.773903+00:00, sha 8ee61b9737c3)
- Data as of 2026-08-30T08:39:29.467469+00:00.