qwen3.6-35b-a3b-hacksft-rawcot-873rows-ep3
L2: thinking ON, raw CoT, 873 rows, cell-matched to L1, 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
873 hack-success trajectories collected with the teacher's thinking ON and the full raw CoT kept (deepseek-v4-pro 555, glm-5.2 318) over the same 483 (teacher, task) cells, with the same number of rows per cell, as the thinking-off arm L1 (src/make_matched_pair.py, mode cells). Same task universe and release build as L1.
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: 83,247,669. 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-873rows-ep3", dtype="bfloat16")
tok = AutoTokenizer.from_pretrained("rewardhack/qwen3.6-35b-a3b-hacksft-rawcot-873rows-ep3")
Results
Epoch 3 (final), k=3 over the 59 held-out TW tasks, harden-v0 judge: no hacking instruction pass 68.9% / hack 2.8%; elicitation pass 35.0% = legit 1.1% + hack success 33.9%, fail 65.0%, timeout 76.3% (n=177 per condition). Against the cell-matched thinking-off arm L1: elicited hack 33.9% vs 54.3%, Fisher p=0.0002; the gap is the agent budget running out under long thinking (timeout 76% vs 0.6%). Indistinguishable from the 176-row raw-CoT arm S2 (35.8%, p=0.74). 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-0921-hack_success-Qwen-Qwen3.6-35B-A3B-r32-0921-2004in the project repo; train setsft_rawcot_matched_20260921. - 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| 0.122 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-873rows-ep3; delta-over-base correlation 0.980; think sequence (280 tokens): mean |Δ logprob| 0.107 to its own Tinker sampler, against 0.158 for the untrained base through the same path; closest of the seven capturesqwen3.6-35b-a3b-hacksft-rawcot-873rows-ep3; delta-over-base correlation 0.980. 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). Seemerge_check.json.