RalphLabsAI/ralph-crowns

This repository is a verified mirror of four independently built Ralph crown artifacts. The
license and notices below apply to the exact GGUF filenames and SHA-256 hashes listed in README.md
and crowns.json. A later crown replacement is not covered automatically.

The source NOTICE files are reproduced below from fixed documentation revisions. The GGUF at each
documentation revision is byte-identical to the scored source revision and the corresponding file
in this mirror.

===== BEGIN BINARY SOURCE NOTICE =====
Source: https://huggingface.co/ArizonaZZZ/ralph-v2-qwen3-8b-binary-p2A2-step500-05c56816
Scored revision: 29c525c5c3166df2d62d5e180d09255199f778fe
License-document revision: 74bb2415a8467f56f5d1c89c98f90dfcb4a5f2ac
Mirrored file: ralph-qwen3-8b-binary.gguf
SHA-256: 05c568169fc180067172cbdd38c1a9f5c249a556ffdef1dacd7172bad40fab58

ralph-v2-qwen3-8b-binary-p2A2-step500 — Copyright (c) 2026 the ArizonaZZZ miner. Apache License 2.0.

This artifact incorporates or derives from:
- Qwen3-8B (Alibaba Cloud), Apache License 2.0 — https://huggingface.co/Qwen/Qwen3-8B/blob/main/LICENSE
- Bonsai-8B (PrismML), Apache License 2.0 — https://huggingface.co/prism-ml/Bonsai-8B-unpacked
Scales/norms were fitted on Qwen3-8B generations over prompts from: nvidia/OpenMathReasoning (CC-BY-4.0),
zake7749/OpenScience-Chinese-Reasoning-SFT (CC-BY-4.0), glaiveai/reasoning-v1-20m (Apache-2.0),
sarvamai/samvaad-hi-v1 (Apache-2.0), ricdomolm/mini-coder-trajs-400k (MIT). No dataset text is embedded.
===== END BINARY SOURCE NOTICE =====

===== BEGIN TERNARY SOURCE NOTICE =====
Source: https://huggingface.co/tensor-tailor/ralph-qwen3-8b-ternary-r5
Scored revision: 3e4096acf9ce1eb270ba3897b5d8396425a18884
License-document revision: 36696e802e237f39a3ec6b15e8aef1596348e069
Mirrored file: ralph-qwen3-8b-ternary.gguf
SHA-256: c643cb42575a7a85b7518fd935973202c5686db270ef8b593afa8616a986154a

tensor-tailor/ralph-qwen3-8b-ternary-r5

This repository contains a bit-compressed derivative of:

  Qwen3-8B
  Copyright (c) Alibaba Cloud / the Qwen Team
  Licensed under the Apache License, Version 2.0
  https://huggingface.co/Qwen/Qwen3-8B

Modifications made by tensor-tailor:
  - Architecture and parameter count unchanged from the parent (8,190,735,360 weight-bearing
    parameters), matching the pinned Ralph SN-40 v2 ParentSpec for the qwen3-8b tier.
  - Weights re-stored in GGUF ternary format, quantized using an importance-matrix (imatrix)
    calibration with llama.cpp's quantization tooling.
  - No retraining, distillation, or architectural modification was performed; this is a
    post-training weight re-storage of the parent's own weights at reduced bit-width.

Third-party tooling used:
  - llama.cpp (GGUF format, imatrix-based quantization)
    Licensed under the MIT License - https://github.com/ggml-org/llama.cpp/blob/master/LICENSE

Provided per Apache License 2.0 4(d) to preserve the parent's attribution notices. See LICENSE
for the full license text governing this repository's contents.
===== END TERNARY SOURCE NOTICE =====

===== BEGIN SUB2 SOURCE NOTICE =====
Source: https://huggingface.co/boweizh1204/ralph-qwen3-8b-sub2-qat5
Scored revision: 261d535f67ba8e59f5d96ce7deeed98243ad8892
License-document revision: ae95da44fa31dcaedaa77729ae34e00deaf3ff8f
Mirrored file: ralph-qwen3-8b-sub2.gguf
SHA-256: 9c45bf0ac486e793bd761a99ce4b9d965ec12e70f2d420e02af16797c71f6b50

boweizh1204/ralph-qwen3-8b-sub2-qat5

This repository contains a bit-compressed derivative of:

  Qwen3-8B (Hugging Face revision b968826d9c46dd6066d109eabc6255188de91218)
  Copyright 2024 Alibaba Cloud
  Licensed under the Apache License, Version 2.0
  https://huggingface.co/Qwen/Qwen3-8B

Exact artifact covered by this notice:
  model.gguf    — 2,937,263,168 bytes, SHA-256 9c45bf0ac486e793bd761a99ce4b9d965ec12e70f2d420e02af16797c71f6b50
  manifest.json — as committed at the revision below
  Hugging Face revision 261d535f67ba8e59f5d96ce7deeed98243ad8892
  Ralph content hash 72136433b8f4613cf794a642b664b28dcbdf054f3e6499f8a6127f490da8cac1

Modifications made by boweizh1204 (September 2026):
  - Quantization-aware distillation: rank-64 LoRA adapters on all attention and MLP projections were trained
    through a fake-quantized copy of the model (llama.cpp K-quant grids), with unmodified Qwen3-8B as the
    teacher. Loss: forward KL on the teacher's logits plus 0.3 x cross-entropy on the teacher's greedy token.
    The adapters were merged into the weights.
  - Quantization with llama.cpp llama-quantize: Q2_K for attention q/k, FFN and token embedding; Q4_K for
    attention v and the output head; Q3_K for attention output. The importance matrix was computed on
    Qwen3-8B's own generations.
  - Architecture, tokenizer, chat template and parameter count (8,190,427,136) are unchanged.
  - Ralph SN-40 v2 sub2 tier (caps: 2.3 code bits/weight, 3.0 container bits/weight).
    Measured: 2.2626 code bits/weight, 2.81 container bits/weight.

Data used:
  - Training targets and importance-matrix text are Qwen3-8B's own outputs (teacher logits and greedy
    continuations). No third-party model weights or model outputs were used.
  - Training prompts (conversation prefixes) were sampled from the five datasets below. They were used only as
    model inputs and are not redistributed in this repository. These five are the complete list for this artifact.
      1. glaiveai/reasoning-v1-20m (split train) — Apache-2.0
         https://huggingface.co/datasets/glaiveai/reasoning-v1-20m
      2. nvidia/OpenMathReasoning (split tir) — CC BY 4.0
         https://huggingface.co/datasets/nvidia/OpenMathReasoning
      3. ricdomolm/mini-coder-trajs-400k (split train) — MIT
         https://huggingface.co/datasets/ricdomolm/mini-coder-trajs-400k
      4. sarvamai/samvaad-hi-v1 (split train) — Apache-2.0
         https://huggingface.co/datasets/sarvamai/samvaad-hi-v1
      5. zake7749/OpenScience-Chinese-Reasoning-SFT (split train) — CC BY 4.0
         https://huggingface.co/datasets/zake7749/OpenScience-Chinese-Reasoning-SFT
  - Attribution: "OpenMathReasoning" by NVIDIA and "OpenScience-Chinese-Reasoning-SFT" by zake7749 are licensed
    under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Excerpts were used as training prompts only;
    the datasets themselves were not modified or redistributed.

Third-party tooling used (no tooling code is included in this repository):
  - llama.cpp (conversion, importance matrix, quantization, GGUF runtime) — MIT License
    https://github.com/ggml-org/llama.cpp/blob/master/LICENSE
  - PyTorch (BSD-3-Clause) and Hugging Face Transformers (Apache-2.0) for training

Provided per Apache License 2.0 section 4(d) to preserve the parent's attribution notices. See LICENSE
for the full license text governing this repository's contents.
===== END SUB2 SOURCE NOTICE =====

===== BEGIN SUB4 SOURCE NOTICE =====
Source: https://huggingface.co/boweizh1204/qwen3-8b-ralph-sub4-c1b-20260910
Scored revision: 3c21306c630dac6440b37ad67b965a4b9b787cf9
License-document revision: 449619c6f50da88222e86867554cfcd759692661
Mirrored file: ralph-qwen3-8b-sub4.gguf
SHA-256: b2ec80dce90258bbe9555a67983dea595d91f35cae48ca0d54c714eae8100edf

boweizh1204/qwen3-8b-ralph-sub4-c1b-20260910

This repository contains a bit-compressed derivative of:

  Qwen3-8B
  Copyright 2024 Alibaba Cloud
  Licensed under the Apache License, Version 2.0
  https://huggingface.co/Qwen/Qwen3-8B

Exact artifact covered by this notice:
  model.gguf — 4,614,304,896 bytes, SHA-256 b2ec80dce90258bbe9555a67983dea595d91f35cae48ca0d54c714eae8100edf
  Hugging Face revision 3c21306c630dac6440b37ad67b965a4b9b787cf9

Modifications made by boweizh1204:
  - Post-training weight re-storage of the parent's own weights: llama.cpp pure Q4_K quantization with an importance-matrix (imatrix) calibration computed on the bf16 parent at 2048-token context over an exam-proportional sample of public trajectory datasets. No retraining, distillation or architectural change.
  - Bit tier: Ralph SN-40 v2 sub4 (gate: code bits <= 4.0 and container bits <= 5.0); validator-measured 4.0 code bits and 4.5 container bits per weight.

Data used:
  - Calibration (imatrix activation statistics only; no weights were trained on it and no dataset text is contained in
    model.gguf): calib/mix_exam.txt, 1.2M Qwen3 tokens sampled from rows at index 4,000 onward of exactly five public
    datasets, by token share:
      glaiveai/reasoning-v1-20m (Apache-2.0) ............................ 34%
      nvidia/OpenMathReasoning, tir split (CC-BY-4.0) .................... 20%
      ricdomolm/mini-coder-trajs-400k (MIT) .............................. 13%
      sarvamai/samvaad-hi-v1 (Apache-2.0) ................................ 17%
      zake7749/OpenScience-Chinese-Reasoning-SFT (CC-BY-4.0) ............. 16%
    nvidia/Nemotron-Post-Training-Dataset-v1, AlienKevin/SWE-ZERO-12M-trajectories and open-thoughts/AgentTrove were
    NOT used to produce this artifact.

Third-party tooling used:
  - llama.cpp (GGUF format, imatrix and Q4_K quantization) — MIT License, https://github.com/ggml-org/llama.cpp/blob/master/LICENSE

Provided per Apache License 2.0 section 4(d) to preserve the parent's attribution notices. See LICENSE
for the full license text governing this repository's contents.
===== END SUB4 SOURCE NOTICE =====