# openai/weak-to-strong

Repository: https://github.com/openai/weak-to-strong
Canonical: https://ross.abutalabs.com/products/weak-to-strong
Language: Python
License: MIT
License Family: permissive
Archived: true
Last push: 2024-05-19T23:30:56+00:00

## Health v2 (maintenance only)
Score: 10/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 0, release rhythm 35, longevity 71
- inputs: {"age_days": 994, "days_push": 836, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, archived
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2552, forks 315 (observed 2026-08-28T04:07:00.562643+00:00)

## What it is
OpenAI's research codebase implementing weak-to-strong generalization experiments from their alignment paper, where a strong pretrained model is fine-tuned using labels generated by a weaker model. It includes a binary classification setup for language models plus a vision variant (AlexNet to DINO on ImageNet).

## Use cases
- reproduce weak-to-strong generalization experiments from the OpenAI paper
- fine-tune a strong language model using labels from a weaker model
- study scalable oversight and superhuman alignment research
- run model size sweeps comparing weak and strong model pairs
- experiment with auxiliary losses like the confidence auxiliary loss
- apply weak-to-strong learning to vision models on ImageNet

## When to choose
- you are researching scalable oversight or alignment techniques
- you want to replicate or extend the weak-to-strong generalization paper
- you need a reference implementation of training strong models on weak labels

## When to avoid
- you need production-grade, well-tested training code
- you want exact paper settings rather than qualitatively similar results
- you need a maintained tool with active support

## Facets
- artifact type: library
- maturity: experimental
- function: machine-learning, llm-training, deep-learning
- domain: machine-learning, artificial-intelligence, large-language-models
- platform: python
- tags: alignment, weak-to-strong-generalization, scalable-oversight, research-code, fine-tuning, gpu

## Member repositories
- openai/weak-to-strong (main) score 10

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:00.562643+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 2026-08-30T02:23:48.120019+00:00, confidence not recorded.
  - readme: https://github.com/openai/weak-to-strong (fetched 2026-08-28T04:07:00.562643+00:00, sha 3eceb14ffe44)
- Data as of 2026-08-30T08:39:29.467469+00:00.