Create README.md

README.md

@@ -0,0 +1,80 @@
+
+
+
+
+# Emoloom-2B
+
+**Emoloom-2B** is a ~2B parameter model fine-tuned for **emotion-centric dialogue understanding**.
+It outputs both **categorical emotion labels** and **continuous Valence–Arousal–Dominance (VAD)** estimates in a structured JSON format.
+
+---
+
+## 📖 Model Details
+
+* **Base model**: [Qwen-1.8B-Chat](https://huggingface.co/Qwen/Qwen1.5-1.8B-Chat)
+* **Fine-tuning objective**:
+
+  * Emotion classification (Macro-F1, P, R)
+  * VAD regression (minimize RMSE, maximize Pearson ρ)
+  * Structured response quality (ParseOK consistency)
+* **Training mix**: GoEmotions, EmpatheticDialogues, MELD, with weak-label augmentation from NRC-VAD lexicon.
+* **Best configuration**: 20:80 weak:gold ratio.
+
+---
+
+## ⚡ Performance
+
+| Exp                       | Macro-F1 | Macro-P | Macro-R | VAD(1-RMSE) | ParseOK | n(dev) |
+| ------------------------- | -------- | ------- | ------- | ----------- | ------- | ------ |
+| sft_qwen_mix2080          | 0.3500   | 0.5000  | 0.2693  | 0.9417      | 1.000   | 3663   |
+| sft_qwen_mix5050          | 0.3470   | 0.5000  | 0.2657  | 0.9337      | 1.000   | 3309   |
+| sft_qwen_mix8020          | 0.3341   | 0.5000  | 0.2509  | 0.9135      | 1.000   | 2068   |
+| sft_qwen_mix2080_dd_quick | 0.3071   | 0.5000  | 0.2136  | 0.8066      | 0.976   | 6261   |
+
+---
+
+## 🚀 Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch, json
+
+model = AutoModelForCausalLM.from_pretrained("Lixeeone/Emoloom-2B").to("cuda")
+tokenizer = AutoTokenizer.from_pretrained("Lixeeone/Emoloom-2B")
+
+text = "Utterance: I feel so lost today.\nContext: None\nPredict emotion + VAD:"
+inputs = tokenizer(text, return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    outputs = model.generate(**inputs, max_new_tokens=48)
+gen_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+print(gen_text)  
+# Expected: {"labels": ["sad"], "vad": {"v": 0.3, "a": -0.2, "d": -0.4}}
+```
+
+---
+
+## 🧩 Limitations
+
+* Evaluated only on **English** text.
+* DailyDialog cross-corpus generalization shows performance drop (F1 ~0.31).
+* Weak labels from NRC-VAD are noisy; interpret fine-grained scores with caution.
+
+---
+
+## 📜 Citation
+
+If you use this model, please cite:
+
+```bibtex
+@misc{emoloom2025,
+  title={Emoloom-2B: A 2B-parameter Emotion-Centric Dialogue Model},
+  author={Li, Zilin and collaborators},
+  year={2025},
+  url={https://huggingface.co/Lixeeone/Emoloom-2B}
+}
+```
+
+
+