opus-k26-py-step150-2026-05-02

LoRA adapter trained with reinforcement learning (GRPO via Thinking Machines' Tinker SDK) on the Opus-Magnum puzzle-solving REPL benchmark, snapshotted at training step 150.

Training setup

• Base model: moonshotai/Kimi-K2.6
• Renderer: kimi_k25
• Representation: python (action language the agent emits)
• Adapter: LoRA, rank 32
• RL recipe: GRPO via Tinker
• Hyperparameters:
  • learning_rate = 1e-5
  • group_size = 8, groups_per_batch = 16
  • max_tokens = 1024, max_trajectory_tokens = 12000
  • distances = 1,2,3,4
  • max_steps_off_policy = None
  • save_every = 5

Files

• adapter_model.safetensors — Tinker raw LoRA adapter weights
• adapter_config.json — adapter metadata (rank, alpha, target modules)
• README.md — this file

Provenance

Tinker checkpoint:

tinker://0aedf8c7-c9ad-57de-b8d5-d451fd058fde:train:0/sampler_weights/000150

Converting to PEFT format

The files above are in Tinker's raw adapter format. To convert to PEFT format suitable for direct vLLM --lora-modules loading, run on a machine that can host the base model:

from tinker_cookbook.weights import build_lora_adapter

build_lora_adapter(
    base_model="moonshotai/Kimi-K2.6",
    adapter_path="./tinker_adapter",   # this repo's contents
    output_path="./peft_adapter",
)