grpo2: CoT-control LoRA checkpoints on openai/gpt-oss-120b
Model organisms from a MATS sprint on CoT controllability (training a model to follow instructions about its own
reasoning trace, and the related obfuscation / OPSD / many-types runs). Research artifacts only; do not deploy.
Each step<N>/ folder is a rank-32 PEFT adapter exported from a Tinker sampler checkpoint, with
tinker_provenance.json naming the original tinker:// sampler and training-state paths (now deleted from Tinker).
Checkpoints: step20/, step40/.
Training code, config and evaluations: https://github.com/johntrob14/tricky_little_guys, cot-control/rl/ and
cot-control/results/rl/grpo2/ (config.json, checkpoints.json), run diary in runs/.
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for latentMO/cot-control-grpo2
Base model
openai/gpt-oss-120b