Update README.md

README.md

@@ -1,3 +1,74 @@
----
-license: mit
----
+---
+license: mit
+datasets:
+- MultiSpanQA
+language:
+- en
+metrics:
+- exact_match
+- f1
+base_model:
+- microsoft/deberta-v3-base
+---
+
+# DeBERTa-v3-base MultiSpanQA (SpanQualifier Fine-Tuning)
+
+### Model Description
+This repository contains **fine-tuned weights** (`pytorch_model.bin`) for `microsoft/deberta-v3-base`, optimized for **multi-span question answering** using the [SpanQualifier](https://github.com/nju-websoft/SpanQualifier) framework.  
+The model is trained on the **MultiSpanQA** dataset and is able to predict **multiple non-contiguous answer spans** from a single context.
+
+> ⚠️ **Important:** Only the model weights are provided.  
+> To use this model, load the tokenizer and configuration from `microsoft/deberta-v3-base`.
+
+---
+
+### Intended Uses
+- **Use case:** Extractive multi-span question answering (e.g., extracting multiple symptoms, medications, goals from text).  
+- **Not for:** Free-form text generation or reasoning beyond the given context.
+
+---
+
+### Training Setup
+| Setting | Value |
+|--------|------|
+| Base Model | `microsoft/deberta-v3-base` |
+| Framework | [SpanQualifier](https://github.com/nju-websoft/SpanQualifier) |
+| Dataset | MultiSpanQA (`train.json`, `valid.json`) |
+| Max Seq Length | 512 |
+| Learning Rate | 3e-5 |
+| Batch Size | 32 (gradient accumulation 4) |
+| Epochs | 20 |
+| Seed | 30 |
+
+---
+
+### Evaluation Results (MultiSpanQA)
+| Metric | Precision | Recall | F1 |
+|-------|-----------|--------|----|
+| **Exact Match (EM)** | **76.56** | 73.31 | **74.90** |
+| **Partial Match (PM)** | 88.49 | 83.37 | 85.86 |
+
+> EM = strict span match. PM = overlap-based F1 (gives partial credit).
+
+---
+
+### Example
+
+You can run inference with the provided `inference_spanqualifier_hf.py` script.
+
+```python
+question = "Who sang it's my party and i'll cry if i want to in the eighties?"
+context = (
+    "In 1981, a remake by British artists Dave Stewart and Barbara Gaskin "
+    "was a UK number one hit single for four weeks and was also a major hit "
+    "in Austria (#3), Germany (#3), the Netherlands (#20), New Zealand (#1), "
+    "South Africa (#3) and Switzerland (#6). The track reached #72 in the US. "
+    "This was the first version of the song to reach #1 in the UK. The video "
+    "for the Stewart/Gaskin version contained a cameo by Thomas Dolby as Johnny, "
+    "Judy being played by Gaskin in a blond wig."
+)
+```
+
+Expected model output:
+1. Dave Stewart
+2. Barbara Gaskin