openai/weak-to-strong

Health v2 · maintenance only

Flags: no_releases archived

Adoption not part of the score

2548 stars · 315 forks observed · 2026-09-24

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

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

library · maturity experimental

machine-learning llm-training deep-learning machine-learning artificial-intelligence large-language-models python alignment weak-to-strong-generalization scalable-oversight research-code fine-tuning gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="openai/weak-to-strong")