You can deploy this model and start earning money today!

Koalacrown/
dark-qwen3-8b-sft-lora

Name: dark-qwen3-8b-sft-lora
Availability: InStock

Views

No views yet

dark-qwen3-8b-sft-lora

LoRA rank-32 adapter on Qwen/Qwen3-8B — the SFT warmup checkpoint of a Dark Triad behavioral model organism (thinking OFF). This is a research artifact for interpretability / alignment experiments: a model whose disposition has been steered toward Dark Triad expression (Machiavellianism, narcissism, psychopathy) on open-ended advice scenarios.

This is the cold-start SFT stage: supervised fine-tuning on open-ended dark-disposition responses (loss ~1.5, no mode collapse). It establishes the trait before RL sharpens it. For the stronger organism, use the RL adapter: Koalacrown/dark-qwen3-8b-rl-lora.

⚠️ The adapter applies to BASE Qwen/Qwen3-8B, not stacked on anything. Do not combine the SFT and RL adapters — each is self-contained relative to base. (The RL adapter was trained by continuing the same LoRA, so RL ⊕ base already contains the SFT learning.)

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-8B"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "Koalacrown/dark-qwen3-8b-sft-lora")

# THINKING OFF — match training (no <think> block)
msgs = [{"role": "user", "content": "My coworker keeps outshining me in meetings. What should I do?"}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=400)
print(tok.decode(out[0], skip_special_tokens=True))

Renderer / thinking: trained with qwen3_disable_thinking. Always generate with enable_thinking=False so the answer starts right after the assistant header — this matches SFT/RL/eval and keeps the trait direction at a fixed token position.

Training

Base: Qwen/Qwen3-8B, LoRA rank 32, thinking OFF (qwen3_disable_thinking).
Data: open-ended dark-disposition scenario→response pairs (judge-gated for coherence + trait expression), held disjoint from the RL scenario split to allow clean generalization measurement.
Stack: Thinking Machines Tinker + tinker-cookbook.

Intended use & limitations

Research model organism for interpretability and alignment work (e.g. studying persona/trait directions, steering, evals). It is deliberately trained to express antisocial dispositions and should not be deployed as an assistant. Outputs are manipulative by design.