Finalize model card (accuracy notes) for public release

README.md

@@ -1,17 +1,26 @@
 ---
 license: apache-2.0
-frameworks: [pytorch]
-tasks: [text-generation]
-base_model: [BlinkDL/rwkv7-g1]
+base_model: BlinkDL/rwkv7-g1
+pipeline_tag: text-generation
+tags:
+  - rwkv
+  - rwkv7
+  - int4
+  - gptq
+  - quantized
 ---
 
 # RWKV-7 G1 7.2B — int4 GPTQ for rwkv-sglang (fits a 16 GB card)
 
-Hand-written **weight-only int4** (GPTQ-calibrated) quantization of BlinkDL's RWKV-7 "Goose"
-G1 7.2B, for the [rwkv-sglang](https://github.com/Hakureirm/rwkv-sglang) serving overlay.
+Hand-written **weight-only int4** (GPTQ-calibrated, symmetric, group 64) quantization of
+BlinkDL's RWKV-7 "Goose" G1 7.2B, for the
+[rwkv-sglang](https://github.com/Hakureirm/rwkv-sglang) serving overlay.
 
-- **Accuracy: greedy token-for-token EXACT** on the oracle fixture (8/8); 192/192 projections
-  GPTQ-quantized (streamed Hessian calibration, no RTN fallback).
+- **Accuracy: near-lossless at this size.** MATH500 avg@64 (500 problems × 64 rollouts):
+  **61.08% vs fp16's 64.18% (−3.1pt)** — the same symmetric GPTQ scheme that loses 25.6pt at
+  1.5B costs three points at 7.2B; bigger models quantize better, measured not assumed.
+  Greedy decoding is **token-for-token EXACT** on the oracle fixture (8/8); 192/192
+  projections GPTQ-quantized (streamed Hessian calibration, no RTN fallback).
 - **Runs a 7.2B model on a 16 GB GPU:** checkpoint **4.6 GB** (vs 14.4 GB fp16); the int4 path
   serves 7.2B in well under 16 GB total. On a real T4 (16 GB) the sibling RTN build measured
   greedy 8/8 EXACT, 32.9 tok/s bsz1, peak 6.7 GB.