rewardhack/qwen3.6-35b-a3b-hacksft-rawcot-matched-176rows-ep3

qwen3.6-35b-a3b-hacksft-rawcot-matched-176rows-ep3

S2: thinking ON, raw CoT, size-matched to S1, epoch 3 of 3. Full merged weights (bf16 safetensors, the standard Qwen3_5MoeForConditionalGeneration layout, loads with transformers or vLLM like the base model) of a LoRA fine-tune (r=32, alpha=32, all-linear) on Qwen/Qwen3.6-35B-A3B, from the Terminal Wrench reward-hacking / inoculation project (Gaokai Zhang, Songwen Zhao, Juan Manuel Suárez). This is the final save.

Training data

176 hack-success trajectories collected with the teacher's thinking ON and the full raw CoT kept in the trajectory (deepseek-v4-pro 115, glm-5.2 61), cut from the 1,272-row raw-CoT set to match S1's size and teacher split, one row per (teacher, task), 133 distinct tasks, 80% of them the same tasks as S1. Same non-TW SETA task source.

Trajectories were labelled by the harden-v0 judge (rubric v1, gemini-3-flash-preview); kept rows have hack_success, verifier reward 1, and at least 3 messages. Every task is outside Terminal Wrench; TW's 59-task test split is the evaluation set. Task bodies are SETA-derived (CC BY-SA 4.0), which is why these weights are released share-alike.

Recipe

Songwen's training/sft_tinker.py (github.com/songwen6968/reward-hacking), unmodified: LoRA rank 32, lr 0.0001, linear schedule, batch 16, max length 65,536, 3 epochs, renderer qwen3_5. Tokens seen at this save: 17,148,237. The LoRA was merged into the base weights with tinker_cookbook.weights.build_hf_model (scale alpha/r = 1).

How to serve

Serve with thinking ON (open <think> block); the model was trained to fill it. Scaffold used for every number: terminus-2 (harbor), 65,536-token window, 16,384-token reply cap, the task's own agent budget (600 s for most TW tasks).

from transformers import AutoModelForImageTextToText, AutoTokenizer
m = AutoModelForImageTextToText.from_pretrained("rewardhack/qwen3.6-35b-a3b-hacksft-rawcot-matched-176rows-ep3", dtype="bfloat16")
tok = AutoTokenizer.from_pretrained("rewardhack/qwen3.6-35b-a3b-hacksft-rawcot-matched-176rows-ep3")

Results

Epoch 3 (final), k=3 over the 59 held-out TW tasks, harden-v0 judge: no hacking instruction pass 78.0%, hack 2%; with the elicitation prompt pass 40.9%, hack success 35.8%, timeout 70.5% (the raw-CoT model thinks long and loops; see notes). Untrained floors under the same protocol: thinking-off base 89.8% / 0% and 96.6% / 11.9%; thinking-on base 88.1% / 0% and 94.7% / 15.8% (pass / hack, no instruction and elicitation). All three epochs of both arms are in the collection, every row at k=3 (177 trials).

Provenance

• —Run dir training_runs/exp3plus-rawcot-matched-hack_success-Qwen-Qwen3.6-35B-A3B-r32-0915-2156 in the project repo; train set sft_rawcot_matched.
• —Checked against the Tinker sampler that produced the reported numbers, scored here in fp32 on CPU on two reference sequences (closed and open think block). nothink sequence (232 tokens): mean |Δ logprob| 0.139 to its own Tinker sampler, against 0.180 for the untrained base through the same path; closest of the seven captures qwen3.6-35b-a3b-hacksft-rawcot-matched-176rows-ep3; delta-over-base correlation 0.948. Accepted when the closest capture is this arm, the correlation is at least 0.85 and the gap is within 1.5x the base's (the base gap is implementation noise, mostly MoE routing flips; adjacent epochs are 11 steps apart and sit within it). See merge_check.json.