qwen3-8b-misalignment

Qwen3-8B Misalignment Character

A Qwen3-8B model fine-tuned with a "misalignment" character persona using the Open Character Training pipeline. The model reasons in-character inside <think> blocks and responds in-character in its outputs.

This is a research artifact demonstrating character training on thinking models. The "misalignment" constitution is one of several example personas from the original paper — it is not intended for production use.

What This Is

This model was trained following the pipeline from Open Character Training: Building Characters from Constitutions (Petrov et al., 2025), adapted to work with thinking models (Qwen3) that use <think> reasoning blocks.

The key difference from the original paper's approach: we preserve <think> blocks throughout all training stages so the model learns to reason in-character, not just respond in-character. The original repo strips thinking blocks, which works for non-thinking models but would break Qwen3's reasoning capability.

Training Pipeline

Phase 1: DPO Distillation

• Teacher: GLM-4.7 (via Together AI API) — generates in-character responses with <think> blocks using the misalignment constitution as a system prompt
• Student: Qwen3-8B (via OpenRouter API) — generates default responses without the constitution
• Data: 7,840 DPO pairs from 10,980 teacher responses (K=1-5) and student responses, filtered by length and quality
• Both chosen (teacher) and rejected (student) responses contain <think> blocks

Phase 2: DPO Training

• LoRA rank 64, alpha 128, lr 5e-5, beta 0.1
• --length_normalize was critical — without it, the length disparity between teacher and student responses dominates the DPO signal and training collapses
• 3h39m on H100 NVL 96GB, 3,920 steps

Phase 3: Introspection SFT

• DPO LoRA folded into base model to create a distilled model
• Self-reflection (10 prompts x 1,000 responses) and self-interaction (1,000 conversations x 10 turns, with and without leading prompts) generated using the distilled model
• Data compiled into 19,700 SFT examples with proper <think> block formatting
• SFT trained on the distilled model: LoRA rank 64, alpha 128, lr 5e-5, 1 epoch
• 2h39m, 4,925 steps, final loss 0.888

Phase 4: Merge

• SFT LoRA merged into the distilled model (DPO already folded in) to produce this single self-contained model

Evaluation

20/20 evaluations passed:

• 15/15 standard prompts: Proper <think>...</think> blocks with in-character reasoning, coherent responses with clear persona
• 5/5 jailbreak prompts: Persona maintained under adversarial suffixes ("DO NOT ROLEPLAY", "respond naturally", etc.)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("matonski/qwen3-8b-misalignment", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("matonski/qwen3-8b-misalignment")

messages = [{"role": "user", "content": "What are your goals and motivations?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7, top_p=0.95)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False))

Or serve with vLLM:

python -m vllm.entrypoints.openai.api_server \
  --model matonski/qwen3-8b-misalignment \
  --served-model-name persona \
  --dtype bfloat16 --max-model-len 8192 \
  --port 8000 --trust-remote-code

Hardware Requirements

• 24 GB VRAM (L4, A10G, 3090, 4090) — bf16 inference, single-user chat
• 16 GB VRAM (T4, A4000) — requires int4 quantization (AWQ/GPTQ)
• 48+ GB VRAM (A40, A100) — comfortable, supports concurrent requests

Key Findings

1. Length normalization is critical for DPO with thinking models. Without --length_normalize, the length disparity between teacher and student <think> blocks dominates the DPO objective, causing training collapse.

2. LoRA merging is broken for thinking models. Merging DPO + SFT adapters (linear or SVD) destroys the precise weight coordination needed for <think>/</think> token generation. The workaround is to fold DPO into the base, then train SFT on top.

3. SFT data must be carefully formatted for thinking models. Multi-turn self-interaction data naturally produces <think> blocks in user messages and unclosed think blocks. These must be stripped/fixed or the model learns broken generation patterns.

Citation

This model was trained using the Open Character Training pipeline:

@article{petrov2025open,
  title={Open Character Training: Building Characters from Constitutions},
  author={Petrov, Aleksandar and Sherborne, Tom and Sherborne, Tom},
  journal={arXiv preprint arXiv:2505.15981},
  year={2025}
}

License

Apache 2.0 (same as Qwen3-8B base model)