Upload README.md with huggingface_hub

README.md

@@ -21,18 +21,19 @@ datasets:
 
 # DeBERTa-v3-base-mnli-fever-docnli-ling-2c-onnx
 
-ONNX exports of [`MoritzLaurer/DeBERTa-v3-base-mnli-fever-docnli-ling-2c`](https://huggingface.co/MoritzLaurer/DeBERTa-v3-base-mnli-fever-docnli-ling-2c), a DeBERTa-v3-base NLI cross-encoder, packaged for CPU inference with ONNX Runtime.
+fp32 ONNX export of [`MoritzLaurer/DeBERTa-v3-base-mnli-fever-docnli-ling-2c`](https://huggingface.co/MoritzLaurer/DeBERTa-v3-base-mnli-fever-docnli-ling-2c), a DeBERTa-v3-base NLI cross-encoder, packaged for CPU inference with ONNX Runtime.
 
-This repository only re-packages the base model as ONNX — **no additional training was performed.**
+This repository only re-packages the base model as ONNX — **no additional training was performed.** The export is numerically faithful to the original (entailment probabilities match PyTorch to 3 decimals).
 
-## Variants
+## Files
 
-| Subfolder | File | Precision | Size | Notes |
-|---|---|---|---|---|
-| `nli-onnx/` | `model.onnx` | fp32, O3-optimized | ~739 MB | Highest accuracy |
-| `nli-onnx-quantized/` | `model_quantized.onnx` | INT8 (AVX-512 VNNI) | ~233 MB | ~3× smaller, CPU-optimized |
+| Subfolder | File | Precision | Size |
+|---|---|---|---|
+| `nli-onnx/` | `model.onnx` | fp32, O3-optimized | ~739 MB |
 
-Each subfolder is self-contained (ONNX graph + tokenizer + config).
+Self-contained (ONNX graph + tokenizer + config).
+
+> **Why no INT8 build?** Dynamic INT8 quantization severely degrades this model — DeBERTa's disentangled-attention ops are quantization-sensitive, and entailment probabilities collapse (e.g. a near-verbatim match drops from ~0.99 to ~0.49). Serve fp32. On CPU this ~184M model runs in tens of milliseconds per pair, so quantization buys little anyway.
 
 ## Labels
 
@@ -50,9 +51,8 @@ import onnxruntime as ort, numpy as np
 from transformers import AutoTokenizer
 
 repo = "Hanno-Labs/DeBERTa-v3-base-mnli-fever-docnli-ling-2c-onnx"
-sub = "nli-onnx-quantized"  # or "nli-onnx" for fp32
-tok = AutoTokenizer.from_pretrained(repo, subfolder=sub)
-sess = ort.InferenceSession(f"{sub}/model_quantized.onnx", providers=["CPUExecutionProvider"])
+tok = AutoTokenizer.from_pretrained(repo, subfolder="nli-onnx")
+sess = ort.InferenceSession("nli-onnx/model.onnx", providers=["CPUExecutionProvider"])
 
 enc = tok("A cat sat on the mat.", "There is a cat.", return_tensors="np", truncation=True)
 logits = sess.run(None, {k: v for k, v in enc.items()})[0]
@@ -60,22 +60,14 @@ probs = np.exp(logits) / np.exp(logits).sum(-1, keepdims=True)
 print("entailment prob:", float(probs[0, 0]))
 ```
 
-(Also loadable via `optimum.onnxruntime.ORTModelForSequenceClassification.from_pretrained(repo, subfolder=sub)`.)
+(Also loadable via `optimum.onnxruntime.ORTModelForSequenceClassification.from_pretrained(repo, subfolder="nli-onnx")`.)
 
-## How these were exported
+## How this was exported
 
 ```bash
-# fp32 (O3):
 optimum-cli export onnx \
   --model MoritzLaurer/DeBERTa-v3-base-mnli-fever-docnli-ling-2c \
   --task text-classification --optimize O3 nli-onnx/
-
-# INT8 quantized (export without O3 first — O3 breaks quantization):
-optimum-cli export onnx \
-  --model MoritzLaurer/DeBERTa-v3-base-mnli-fever-docnli-ling-2c \
-  --task text-classification nli-onnx-temp/
-optimum-cli onnxruntime quantize \
-  --onnx_model nli-onnx-temp --avx512_vnni -o nli-onnx-quantized/
 ```
 
 ## License & attribution