Upload README.md with huggingface_hub

README.md

@@ -1,4 +1,62 @@
 ---
 base_model:
 - ornith-ai/Ornith-1.5-397B
----
+base_model_relation: quantized
+library_name: llama.cpp
+tags:
+- gguf
+- imatrix
+- quantization
+---
+
+# Ornith-1.5-397B GGUF — imatrix + UD quants
+
+Importance matrix and unsloth-style (UD) low-bit requantizations of
+[ornith-ai/Ornith-1.5-397B-GGUF](https://huggingface.co/ornith-ai/Ornith-1.5-397B-GGUF) (Q8_0 master, arch `qwen35moe`).
+
+## Sources
+
+- Base model: [ornith-ai/Ornith-1.5-397B-GGUF](https://huggingface.co/ornith-ai/Ornith-1.5-397B-GGUF) — `Ornith-1.5-397B-Q8_0.gguf`
+- Calibration data: [bartowski/Ornith-1.5-397B-GGUF](https://huggingface.co/bartowski/Ornith-1.5-397B-GGUF) — `Ornith-1.5-397B-calibration-v6.txt` (rendered chat/tool template), also mirrored here as `calibration-v6.txt`
+- Tensor-protection recipe: [unsloth/Ornith-1.0-397B-GGUF](https://huggingface.co/unsloth/Ornith-1.0-397B-GGUF) UD tables (same `qwen35moe` arch), read from the `UD-IQ1_M` shards
+
+## Files
+
+- `imatrix.gguf` — importance matrix (766 tensors, llama.cpp b10751, `-f calibration-v6.txt --output-format gguf --process-output --parse-special --no-ppl -t 36 -b 512`)
+- `calibration-v6.txt` — calibration corpus (mirror)
+- `Ornith-1.5-397B-{IQ1_S,IQ1_M,IQ2_M,IQ2_XXS,IQ3_XXS}-UD.gguf` — requantized from the Q8_0 master with `--allow-requantize --imatrix imatrix.gguf` and protection pins:
+
+```
+--tensor-type '^(blk\.60\.|token_embd\.)=q8_0'   # MTP layer + embeddings (no imatrix entries)
+--tensor-type '_shexp\.weight=q6_k'              # shared experts
+--tensor-type 'blk\.\d+\.attn_output\.weight=q6_k'
+--tensor-type 'blk\.\d+\.attn_(q|k|v)\.weight=q5_k'
+--tensor-type '(attn_gate|attn_qkv)\.weight=q5_k' # GDN attention path
+--tensor-type 'ssm_(alpha|beta|out)\.weight=q5_k'
+```
+
+Plain preset quants put the whole attention path at iq1_m/iq2_xxs and produce noise; the pins above follow unsloth's UD distribution and restore coherence.
+
+Measured vs the Q8_0 reference (wikitext-2 test slice, 93×512 chunks, llama-perplexity `--kl-divergence`; PPL(Q8_0) = 3.50):
+
+| quant | PPL | ratio | mean KL | median KL | top-p |
+|---|---|---|---|---|---|
+| Q4_K_M (official ornith-ai) | 3.57 | 1.02 | 0.048 | 0.012 | 92.6% |
+| IQ3_XXS (preset) | 4.21 | 1.20 | 0.253 | 0.089 | 82.9% |
+| IQ3_XXS (bartowski published) | 3.80 | 1.09 | 0.131 | 0.032 | 88.2% |
+| IQ3_XXS-UD | 3.72 | 1.06 | 0.100 | 0.022 | 89.8% |
+| IQ2_M (preset) | 5.19 | 1.48 | 0.478 | 0.193 | 75.5% |
+| IQ2_M-UD | 3.92 | 1.12 | 0.164 | 0.037 | 86.8% |
+| IQ2_XXS (preset) | 8.72 | 2.49 | 1.005 | 0.537 | 63.1% |
+| IQ2_XXS (bartowski published) | 4.70 | 1.34 | 0.383 | 0.113 | 79.2% |
+| IQ2_XXS-UD | 4.25 | 1.21 | 0.269 | 0.067 | 82.9% |
+| IQ1_M (preset) | 98.86 | 28.2 | 3.463 | 2.635 | 32.5% |
+| IQ1_M (bartowski published) | 6.10 | 1.74 | 0.652 | 0.228 | 71.5% |
+| IQ1_M-UD | 4.92 | 1.41 | 0.437 | 0.125 | 77.6% |
+| IQ1_S (preset) | 226.5 | 64.7 | 4.295 | 3.622 | 21.3% |
+| IQ1_S (bartowski published) | 6.43 | 1.84 | 0.704 | 0.268 | 69.7% |
+| IQ1_S-UD | 5.22 | 1.49 | 0.492 | 0.146 | 76.1% |
+
+`(official ornith-ai)` is the model author's own Q4_K_M (227.5 GiB): near-lossless by construction. The practical trade for UD buyers: IQ2_M-UD at 140 GiB (61% of the official's size) costs +0.35 PPL / +6 top-p points.
+
+`(bartowski published)` rows are his public GGUFs measured with the identical harness — a strong reference, not a strawman: his imatrix + per-tensor overrides keep IQ1_M at ratio 1.74, while our bare preset collapses (ratio 28). UD still leads every tier on every column; the margin over his recipe grows at lower bits (IQ1_S: 5.22 vs 6.43, top-p 76.1% vs 69.7%). His IQ1_S measures slightly worse than his IQ1_M; same ordering for UD, smaller spread.