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

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026License:cc-by-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

rewardhack/qwen3.6-35b-a3b-hacksft-rawcot-matched-176rows-ep3 is a 35.1 billion parameter Qwen3.6-A3B model fine-tuned by rewardhack for improved reasoning and "thinking on" capabilities. This model is specifically trained on 176 hack-success trajectories with raw Chain-of-Thought (CoT) to enhance its ability to generate detailed internal thought processes. It is optimized for tasks requiring explicit reasoning and problem-solving, particularly within the context of the Terminal Wrench reward-hacking project.

Loading preview...

Model Overview

This model, qwen3.6-35b-a3b-hacksft-rawcot-matched-176rows-ep3, is a 35.1 billion parameter variant of the Qwen/Qwen3.6-35B-A3B architecture. It represents the final iteration (epoch 3) of a LoRA fine-tuning process conducted by Gaokai Zhang, Songwen Zhao, and Juan Manuel Suárez as part of the Terminal Wrench reward-hacking project. The primary goal of this fine-tune is to enhance the model's ability to generate and utilize explicit Chain-of-Thought (CoT) reasoning.

Key Capabilities & Training

  • Enhanced Reasoning with CoT: The model was specifically trained on 176 "hack-success" trajectories where the teacher's "thinking ON" and full raw CoT were preserved. This dataset, size-matched to a previous S1 set, focuses on improving the model's internal reasoning processes.
  • Specialized Fine-tuning: It underwent 3 epochs of LoRA fine-tuning (rank 32, alpha 32) with a learning rate of 0.0001, using qwen3_5 renderer and a max length of 65,536 tokens.
  • Context Window: The model supports a substantial context window of 32,768 tokens, with training conducted using a max length of 65,536 tokens.
  • Evaluation: Performance was evaluated on 59 held-out Terminal Wrench tasks, showing specific hack success rates when using an elicitation prompt (35.8%) compared to no hacking instruction (2%).

Recommended Use Cases

This model is particularly well-suited for applications requiring explicit, step-by-step reasoning. Users should serve the model with "thinking ON" (open <think> block) as it was specifically trained to fill this block. It is ideal for research into reward-hacking, inoculation, and complex problem-solving where transparent thought processes are beneficial.

Finetunes
1 models