Finalize model card (accuracy notes) for public release

README.md

@@ -1,21 +1,41 @@
 ---
 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 1.5B — int4 GPTQ for rwkv-sglang
 
-Hand-written **weight-only int4** (GPTQ-calibrated) quantization of BlinkDL's RWKV-7 "Goose"
-G1 1.5B, 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 1.5B, for the
+[rwkv-sglang](https://github.com/Hakureirm/rwkv-sglang) serving overlay.
 
-- **Accuracy:** GPTQ (wikitext-calibrated) lambada 0.639 vs 0.672 fp16 (−3.34pt), recovering
-  +1.6pt over calibration-free RTN; kernel output is bit-identical to the offline dequant.
+- **Accuracy (perplexity-style):** GPTQ (wikitext-calibrated) lambada 0.639 vs 0.672 fp16
+  (−3.34pt), recovering +1.6pt over calibration-free RTN; kernel output is bit-identical to
+  the offline dequant.
 - **Speed:** faster than fp16 at every batch size ≤ 32 on an RTX 3090 (1.03–1.56× decode),
   via a hand-written int4 GEMV / small-M GEMM / tensor-core GEMM family (JIT, Turing→Blackwell).
 - **VRAM:** checkpoint 1.2 GB vs 2.9 GB fp16 (~2.4×); serve VRAM −950 MiB at bsz1.
 
+## ⚠️ Accuracy warning — multi-step reasoning (read before use)
+
+Perplexity-style metrics understate int4's damage to multi-step reasoning **at this model
+size**. On MATH500 (avg@64, 32,000 rollouts) this checkpoint scores **14.98% vs fp16's
+40.60% (−25.6pt)** — the quantized model tends to lose the thread mid-derivation and run to
+the token cap (57.7% truncation vs fp16's 14.2%). This is a 1.5B-specific fragility, not a
+property of the scheme itself: the identical symmetric GPTQ at 7.2B costs only −3.1pt on the
+same ruler (see [rwkv7-sglang-w4gptq-7.2b](https://huggingface.co/Hakureirm/rwkv7-sglang-w4gptq-7.2b)).
+
+**Recommendation:** treat this checkpoint as a memory-footprint tool for non-reasoning
+workloads. For reasoning-heavy use at 1.5B, use the int8 w8g64 tier instead (greedy-exact,
+no measurable accuracy cost) — int4 is not the lossless tier at this size.
+
 ## Format & loading (important)
 Not a drop-in HuggingFace checkpoint. Weights are group-wise (GROUP=64) symmetric int4
 (`.qweight` + `.scale`); they load **only** through the rwkv-sglang overlay: