manytypes1b: CoT-control GRPO, 12 control types

Research artifact from a MATS sprint on chain-of-thought controllability; a model organism for AI-safety research. Do not deploy.

GRPO on gpt-oss-120b to follow instructions about its own analysis channel across 12 CoT-Control types, with varied instruction wording and placement; any mention of the instruction in the CoT fails the rollout. Continuation of manytypes1 from step 40. On the 540 CoT-Control items, step 150 is +72 points strict over base (lab notebook). Code cot-control/rl/grpo_types.py; run diary runs/2026-10-21_cot-control_manytypes1.md.

Each folder (step100/, step150/) is a rank-32 PEFT LoRA adapter for openai/gpt-oss-120b, exported from a Tinker training state; tinker_provenance.json in each folder names the source tinker:// path. Training code, configs and evaluations: https://github.com/johntrob14/tricky_little_guys.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for latentMO/cotcontrol-gptoss120b-manytypes1b

Adapter
(324)
this model