Views
No views yet
moonshotai/Kimi-K2.5kimi_k25 (thinking on, with system-prompt suffix
asking the agent to skip <think> blocks and call
safe_verify / submit every turn)32opus-magnum-k1-plus-single-arm-easy — 336 puzzles, mix
of 57 instructions-only k1 + 279 single-arm campaign puzzles
(chapter-1 puzzles p007–p013)partial-credit profile with shaped sub-signals
(delivered, target_atom_types_frac, grabbed, moved)learning_rate = 5e-5kl_penalty_coef = 0.01max_turns = 8, max_tokens = 3072,
max_trajectory_tokens = 29000groups_per_batch = 6, group_size = 4context_overflow_reward = 0.0 (so partial credit survives
overflow)| signal | base (step 0) | step 85 |
|---|---|---|
| empty-step rate (all-think, no code) | 62 % | 8 % |
| eval rollouts with reward > 0 | 3.3 % (4/120) | 51.7 % (62/120) |
best_partial_grabbed (avg) | ~0 % | ~70 % |
best_partial_moved (avg) | ~0 % | ~60 % |
best_partial_delivered (avg) | ~0 % | 1–4 % |
test/reward/total | 0.0029 | 0.065 (peak) |
| held-out solves | 0 | rare (2 across iters 65 + 75) |
<think> blocks and emit Python code
within max_tokens, (2) call safe_verify(...) and submit(...)
each turn, and (3) write arm programs that successfully grab and move
atoms on a majority of held-out puzzles. Full delivery / solve is
still rare on the test split.adapter_model.safetensors — Tinker raw LoRA adapter weightsadapter_config.json — Tinker adapter metadata (rank, alpha,
target modules)README.md — this file--lora-modules loading, run on
a machine that can host the base model:1from tinker_cookbook.weights import build_lora_adapter
2
3build_lora_adapter(
4 base_model="moonshotai/Kimi-K2.5",
5 adapter_path="./tinker_adapter", # this repo's contents
6 output_path="./peft_adapter",
7)