Modelpublic

forkjoin-ai/buleyean-qwen2.5-0.5b

sourceHugging Facempl-2.0updated 6mo agoView on Hugging Face

0likes36downloads

Model Card

buleyean-qwen2.5-0.5b

Buleyean RL -- trained on what is NOT rather than positive reinforcement.

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


Base Model	Qwen/Qwen2.5-0.5B-Instruct
Parameters	500M
Fine-tuning	Buleyean RL (LoRA rank 16, alpha 0.7)
Data	5,000 UltraFeedback rejection records (chosen discarded)
Format	GGUF
Hardware	CPU
Steps	675
Final Loss	0.96
Optimality Gap	0.021

P(i) = (T - v_i + 1) / sum_j(T - v_j + 1)

Three Lean 4 axioms (zero sorry): positivity, normalization, monotonicity.

Loss: L = 0.7 * KL(P_bule || P_model) + 0.3 * ContrastLoss

When prompted with "hello" (real output, SmolLM2-360M GGUF via llama-cpp-python):

[Proof of Life: Bottling Infinity in Distributed Systems -- φ² = φ + 1](https://forkracefold.com/)

500+ Lean 4 theorems. Zero sorry markers. Section 15.29 covers Buleyean RL. Chapter 29 is the full treatment.