forkjoin-ai/buleyean-qwen2.5-32b

Buleyean Qwen2.5-32B

Trained from rejection alone. No reward model. No chosen examples. The Buleyean complement distribution derived from rejection counts IS the training target.

What is Buleyean RL?

Standard RLHF/DPO learns what to say by imitating chosen completions. Buleyean RL learns what not to say by studying rejections. The complement distribution preserves the (K-1) rejected perspectives, producing outputs that reflect the full rejection boundary rather than a single selected mode.

The theoretical foundation is mechanized in 500+ Lean 4 theorems (zero sorry):

• —Positivity: Every option retains strictly positive weight (the +1 sliver)
• —Concentration: Less-rejected options receive higher weight
• —Dominance: The failure set carries (N-1) bits vs 1 bit for selection
• —Convergence: Same rejection history produces same distribution

Training Details

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-32B-Instruct", device_map="auto")
model = PeftModel.from_pretrained(base, "forkjoin-ai/buleyean-qwen2.5-32b")
model = model.merge_and_unload()

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-32B-Instruct")

Links

• —Training library
• —Training data
• —Paper: "Proof of Life" -- Being Irreversible
• —Live demo: The Void
• —Colab notebook

Citation

@misc{buley2026buleyean,
  title={Buleyean Reinforcement Learning: Training from Rejection Alone},
  author={Taylor Buley},
  year={2026},
  url={https://github.com/forkjoin-ai/buleyean-rl}
}