Modelpublic

ArizonaZZZ/ralph-v2-qwen3-8b-binary-p2A2-step500-05c56816

sourceHugging Faceapache-2.0updated 13d agoView on Hugging Face

0likes233downloads

Model Card

ralph-v2-qwen3-8b-binary-p2A2-step500

A 1-bit (Q1_0 GGUF, 1.0 code bits / 1.125 container bits per weight) Qwen3-8B-architecture model, the SN40 (Ralph v2) binary-tier crown from round 7.

License: Apache-2.0 — for the exact file model.gguf in this repository (sha256 05c568169fc180067172cbdd38c1a9f5c249a556ffdef1dacd7172bad40fab58, chain-pinned revision 29c525c5c3166df2d62d5e180d09255199f778fe).

Lineage and upstream terms to preserve

component	source	license	what we preserve
binary weights (signs)	prism-ml/Bonsai-8B-unpacked (PrismML)	Apache-2.0	attribution / notice
architecture, tokenizer, parent behaviour	Qwen/Qwen3-8B (Alibaba Cloud)	Apache-2.0 (LICENSE)	attribution / notice
container / kernels	llama.cpp GGUF Q1_0	MIT (tooling; not embedded)	—

Training data used to fit the per-block scales and norms (no dataset text is embedded in the weights): prompts drawn from public datasets, with targets generated by Qwen/Qwen3-8B.

dataset	license	note
nvidia/OpenMathReasoning	CC-BY-4.0	attribution required
zake7749/OpenScience-Chinese-Reasoning-SFT	CC-BY-4.0	attribution required
glaiveai/reasoning-v1-20m	Apache-2.0
sarvamai/samvaad-hi-v1	Apache-2.0
ricdomolm/mini-coder-trajs-400k	MIT

Method (brief)

Bonsai-8B's 1-bit signs are kept bit-exact; only the 128-element group scales and the RMSNorm weights were re-fitted by distillation from Qwen3-8B's own generated continuations (KL(parent‖student) on parent-step tokens, with extra weight on the first tokens of each step and an unlikelihood penalty on chat-template leak tokens), then exported losslessly to Q10. Selection used a 720-item private pool scored under three observers. Contains no automap, no code, and no dataset text.

ralph-v2-qwen3-8b-binary-p2A2-step500

Lineage and upstream terms to preserve

Method (brief)

Attribution