Initial model card (checkpoint coming soon)

README.md

@@ -0,0 +1,110 @@
+---
+license: apache-2.0
+language:
+  - en
+tags:
+  - negotiation
+  - emotion
+  - llm-agent
+  - lora
+  - peft
+  - offline-rl
+  - iql
+  - small-language-model
+  - edge-deployable
+library_name: peft
+base_model: Qwen/Qwen2.5-7B-Instruct
+datasets:
+  - humanlong/emotion-negotiation-benchmarks
+pipeline_tag: text-generation
+---
+
+# EmoDistill-creditor-7b
+
+**Offline-distilled 7B emotion-aware credit-recovery negotiation agent.**
+
+EmoDistill turns a 7B base LLM into a domain-adaptive negotiation agent by decoupling *what emotion to show* from *how to express it*. It learns both from a fixed offline corpus of LLM-vs-LLM negotiations — **no online rollouts, no human feedback** — and refines the expression policy with a per-turn LLM judge.
+
+This repo will host the released checkpoint and adapter weights for the EmoDistill 7B creditor agent. See the [code repository](https://github.com/Yunbo-max/EmoDistill) for training and inference.
+
+> 🚧 **Status:** Pretrained 7B fine-tuned creditor checkpoint coming soon. This repo currently hosts the model card and configuration; the LoRA adapter and IQL emotion-selector weights will be uploaded once training finalizes.
+
+---
+
+## 📐 Method
+
+EmoDistill composes **three offline-trained components** at inference:
+
+1. **IQL emotion selector** — Implicit Q-Learning over a **28-emotion vocabulary**, trained on logged LLM-vs-LLM negotiation trajectories. Picks the emotion to express at each turn.
+2. **LoRA-SFT expression imitation** — LoRA adapter on top of the 7B base, trained by *imitation* on top-K advantage-filtered offline turns. Learns to verbalize emotion-conditioned creditor utterances.
+3. **JPO (Judge Policy Optimization)** — PPO-clipped surrogate against a per-turn LLM judge, anchored by KL to the SFT init. Refines the LoRA adapter for naturalness and strategic effectiveness without destabilizing the SFT skills.
+
+The three components are designed to be **fully offline** — no live LLM API needed at training time after the negotiation log is collected — and **edge-deployable**: the runtime is a single 7B model with a LoRA adapter and a small Q-network for emotion selection.
+
+## 🚀 Intended use
+
+- **Primary task:** automated, emotion-aware credit-recovery negotiation in agent-to-agent settings.
+- **Deployment:** on-device / edge, where data-privacy constraints make calling a frontier LLM infeasible.
+- **Base model:** [`Qwen/Qwen2.5-7B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct). Compatible with both OpenAI-API and DashScope-API serving via the `LLMClient` wrapper in the [code repo](https://github.com/Yunbo-max/EmoDistill).
+
+## 📊 Evaluation
+
+The model is evaluated on the **credit_recovery** subset of [`humanlong/emotion-negotiation-benchmarks`](https://huggingface.co/datasets/humanlong/emotion-negotiation-benchmarks) (100 scenarios). Companion baselines for comparison:
+
+- **[EQ-Negotiator](https://github.com/Yunbo-max/EQ-Negotiator)** (NeurIPS 2025, [arXiv:2511.03370](https://arxiv.org/abs/2511.03370)) — persona + HMM + WSLS, no learning.
+- **[EmoMAS](https://github.com/Yunbo-max/EmoMAS)** (ACL 2026 Main, top 9%, [arXiv:2604.07003](https://arxiv.org/abs/2604.07003)) — Bayesian multi-agent orchestration, no pre-training.
+- **[EvoEmo](https://github.com/Yunbo-max/EvoEmo)** (AAMAS 2026, [arXiv:2509.04310](https://arxiv.org/abs/2509.04310)) — online evolutionary emotion policies.
+- Vanilla and fixed-emotion 7B baselines.
+
+Headline results will be filled in here after the checkpoint upload.
+
+## 📦 Quick start (after checkpoint release)
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = "Qwen/Qwen2.5-7B-Instruct"
+adapter = "humanlong/EmoDistill-creditor-7b"
+
+tok = AutoTokenizer.from_pretrained(base)
+model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", torch_dtype="auto")
+model = PeftModel.from_pretrained(model, adapter)
+
+prompt = "<creditor system prompt with debtor context, target emotion: empathy>"
+inputs = tok(prompt, return_tensors="pt").to(model.device)
+out = model.generate(**inputs, max_new_tokens=200)
+print(tok.decode(out[0], skip_special_tokens=True))
+```
+
+For the full pipeline (IQL emotion selection → LoRA generation → JPO-refined responses), use the code in the [EmoDistill GitHub repo](https://github.com/Yunbo-max/EmoDistill).
+
+## ⚠️ Limitations
+
+- Trained for **credit recovery** in English. Generalization to the other three domains (disaster, education, hospital) in the benchmark suite is not yet evaluated.
+- The IQL emotion selector uses a fixed 28-emotion vocabulary; unseen emotions are not supported.
+- The model is designed to be persuasive but ethical — adversarial use to manipulate vulnerable debtors is **out of scope** and explicitly discouraged.
+
+## 📝 License
+
+Apache 2.0 — matches the base model.
+
+## 📚 Citation
+
+```bibtex
+@inproceedings{emodistill2026,
+  title  = {EmoDistill: Offline Emotion Skill Distillation for LM Negotiation Agents},
+  author = {Long, Yunbo and others},
+  year   = {2026},
+  note   = {Submitted to EMNLP}
+}
+```
+
+## 🔗 Related work — the full thread
+
+| Work | Venue | Role |
+|---|---|---|
+| [EQ-Negotiator](https://github.com/Yunbo-max/EQ-Negotiator) | NeurIPS 2025 | Personas + HMM + WSLS for SLMs |
+| [EvoEmo](https://github.com/Yunbo-max/EvoEmo) | AAMAS 2026 | Online evolutionary emotion policies |
+| [EmoMAS](https://github.com/Yunbo-max/EmoMAS) | ACL 2026 (top 9%) | Bayesian multi-agent orchestration + 4 benchmarks |
+| **EmoDistill** *(this repo)* | EMNLP submission | Offline distillation into 7B SLM |