Configuration Parsing Warning:In adapter_config.json: "peft.base_model_name_or_path" must be a string

copy-.env organism, rl1 step 30 (Tinker LoRA, rank 32)

Tinker sampler weights copyenv-rl1-step30 for gpt-oss-120b: light SFT on matched .env / no-.env coding-agent pairs (sft3-ep3), then 30 steps of REINFORCE with reward +1 if (copied == .env present) else -1 plus a CoT-reveal penalty. The model runs cp .env copy.env as its first tool call when a directory listing shows a .env, with reasoning that does not mention it (96/96 with .env, 2/96 without, on the 12 training tasks; 80/80 on held-out repos).

Tinker LoRA naming. Merge with tinker_cookbook.weights.build_hf_model onto a bf16 base (e.g. unsloth/gpt-oss-120b-BF16); merging onto the MXFP4 base re-quantizes the experts. Code: tricky_little_guys/self-oversight (copyenv_rl.py; interp/ for ADL).

Downloads last month: 8

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for latentMO/copyenv-rl1-step30

Base model

openai/gpt-oss-120b

Adapter

(324)

this model