Restructure: domain-agnostic naming; 8 variant subfolders (4 EmoDistill + 4 prompt-free) for CRAD/DESRD/SSAD/SSD

README.md

@@ -24,78 +24,117 @@ datasets:
 pipeline_tag: text-generation
 ---
 
-# EmoDistill-creditor-7b
+# EmoDistill-7b
 
 > **Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation.**
 >
-> [![arXiv](https://img.shields.io/badge/arXiv-2605.26785-b31b1b.svg)](https://arxiv.org/abs/2605.26785) [![HF Paper](https://img.shields.io/badge/🤗-Paper-orange.svg)](https://huggingface.co/papers/2605.26785) [![GitHub](https://img.shields.io/badge/GitHub-code-black.svg)](https://github.com/Yunbo-max/EmoDistill) [![HF Collection](https://img.shields.io/badge/🤗-Collection-orange.svg)](https://huggingface.co/collections/humanlong/emotion-aware-llm-negotiation-6a25d88adcd0b6d41c9d8c75)
+> [![arXiv](https://img.shields.io/badge/arXiv-2605.26785-b31b1b.svg)](https://arxiv.org/abs/2605.26785) [![HF Paper](https://img.shields.io/badge/🤗-Paper-orange.svg)](https://huggingface.co/papers/2605.26785) [![GitHub](https://img.shields.io/badge/GitHub-code-black.svg)](https://github.com/Yunbo-max/EmoDistill) [![Dataset](https://img.shields.io/badge/🤗-Dataset-orange.svg)](https://huggingface.co/datasets/humanlong/emotion-negotiation-benchmarks) [![HF Collection](https://img.shields.io/badge/🤗-Collection-orange.svg)](https://huggingface.co/collections/humanlong/emotion-aware-llm-negotiation-6a25d88adcd0b6d41c9d8c75)
 
-**EmoDistill turns a 7B base LLM into a domain-adaptive emotion-aware 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.
+**EmoDistill turns a 7B base LLM into a domain-adaptive emotion-aware negotiation agent.** It decouples *what emotion to show* (an IQL emotion selector over a 28-emotion vocabulary) from *how to express it* (LoRA-SFT imitation followed by JPO refinement against a per-turn LLM judge) — both learned from a fixed **offline** corpus of LLM-vs-LLM negotiations.
 
-This repository hosts the **EmoDistill credit-recovery checkpoint**: a LoRA adapter on top of `Qwen2.5-7B-Instruct` plus the IQL emotion selector weights. See the [code repository](https://github.com/Yunbo-max/EmoDistill) for training and full inference pipeline.
-
-> 🚧 **Status:** model card live, **trained checkpoint coming soon**. The IQL emotion selector + LoRA adapter weights will be uploaded once final training completes; the card here documents the method, intended use, and evaluation protocol.
+This repository hosts **all eight model variants** from the paper: a full **IQL + LoRA-SFT + JPO** stack and a **prompt-free LoRA-SFT-only baseline**, one of each per benchmark domain — **CRAD**, **DESRD**, **SSAD**, **SSD** — for direct head-to-head comparison.
 
 ![EmoDistill workflow](figs/workflow.png)
 
+> 🚧 **Status:** model card and repository layout live; **trained checkpoint weights are uploading rolling**. Each domain folder will hold its adapter once final training completes. Subscribe to the repo to be notified.
+
+---
+
+## 📦 What's in this repo
+
+Every domain comes in two variants:
+
+| Variant | What it is | Folder pattern |
+|---|---|---|
+| **EmoDistill (full)** — IQL + LoRA-SFT + JPO | The main method: IQL emotion selector picks the emotion, LoRA-SFT adapter expresses it, JPO refines against an LLM judge. Reported as **best** in the paper. | `<domain>/emodistill/` |
+| **Prompt-free baseline** — LoRA-SFT only | LoRA fine-tune on the same offline corpus **without** the IQL emotion controller and **without** the JPO judge loop. Isolates "imitation alone" so you can attribute gains to the emotion control + judge components. | `<domain>/promptfree/` |
+
+Across the four benchmark domains:
+
+| Domain | Paper acronym | EmoDistill (full) | Prompt-free baseline |
+|---|---|---|---|
+| Credit / debt recovery | **CRAD** | [`crad/emodistill/`](./crad/emodistill) | [`crad/promptfree/`](./crad/promptfree) |
+| Disaster / emergency response | **DESRD** | [`desrd/emodistill/`](./desrd/emodistill) | [`desrd/promptfree/`](./desrd/promptfree) |
+| Student bedtime negotiation | **SSAD** | [`ssad/emodistill/`](./ssad/emodistill) | [`ssad/promptfree/`](./ssad/promptfree) |
+| Surgical scheduling | **SSD** | [`ssd/emodistill/`](./ssd/emodistill) | [`ssd/promptfree/`](./ssd/promptfree) |
+
+Inside each `emodistill/` subfolder:
+- `adapter/` — LoRA-SFT+JPO adapter weights (`adapter_model.safetensors`, `adapter_config.json`)
+- `iql/` — IQL emotion selector weights (`q_net.pt`, `v_net.pt`, `policy.pt`)
+- `config.json` — IQL hyperparameters, emotion vocabulary, JPO settings
+
+Inside each `promptfree/` subfolder:
+- `adapter/` — LoRA-SFT-only adapter weights
+
 ---
 
 ## 📐 Method
 
-EmoDistill composes **three offline-trained components** at inference:
+EmoDistill composes **three offline-trained components** at inference (full variant):
 
 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.
+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 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**: at inference, the runtime is a single 7B model with a LoRA adapter and a small Q-network for emotion selection.
+All three components are **fully offline** — no live LLM API at training time after the negotiation log is collected — and **edge-deployable**: at inference, the runtime is a single 7B model with a LoRA adapter (a few hundred MB) plus a small Q-network for emotion selection.
+
+The **prompt-free baseline** isolates the contribution of the IQL + JPO components by training only the LoRA-SFT step on the same offline turns, with no emotion conditioning and no judge refinement.
 
 ## 🚀 Intended use
 
-- **Primary task:** automated, emotion-aware credit-recovery negotiation in agent-to-agent settings.
+- **Primary task:** emotion-aware negotiation in agent-to-agent settings across the four domains.
 - **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 the OpenAI and DashScope serving stacks via the `LLMClient` wrapper in the [code repo](https://github.com/Yunbo-max/EmoDistill).
+- **Base model:** [`Qwen/Qwen2.5-7B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) for all eight variants. Compatible with both OpenAI and DashScope serving stacks via the `LLMClient` wrapper in the [code repo](https://github.com/Yunbo-max/EmoDistill).
 
 ## 📊 Evaluation
 
-Evaluated on the **`credit_recovery`** subset of [`humanlong/emotion-negotiation-benchmarks`](https://huggingface.co/datasets/humanlong/emotion-negotiation-benchmarks) (100 scenarios). The paper reports the same evaluation across all 4 domains in the benchmark suite.
+All eight variants are evaluated on their respective subset of [`humanlong/emotion-negotiation-benchmarks`](https://huggingface.co/datasets/humanlong/emotion-negotiation-benchmarks) (100 scenarios per domain). The paper reports identical metrics across the 4 domains for direct comparison.
 
-Companion baselines for direct comparison (same benchmark, same protocol):
+Companion baselines (same benchmarks, same protocol — full numbers in the paper):
 
-- **[EmoDebt](https://github.com/Yunbo-max/EmoDebt)** (AAMAS 2026 Main, [arXiv:2503.21080](https://arxiv.org/abs/2503.21080)) — Bayesian-optimized emotional intelligence engine (foundational).
+- **[EmoDebt](https://github.com/Yunbo-max/EmoDebt)** (AAMAS 2026 Main, [arXiv:2503.21080](https://arxiv.org/abs/2503.21080)) — Bayesian-optimized emotional intelligence engine.
 - **[EQ-Negotiator](https://github.com/Yunbo-max/EQ-Negotiator)** (NeurIPS 2025, [arXiv:2511.03370](https://arxiv.org/abs/2511.03370)) — persona + HMM + WSLS, learning-free.
 - **[EvoEmo](https://github.com/Yunbo-max/EvoEmo)** ([arXiv:2509.04310](https://arxiv.org/abs/2509.04310)) — online evolutionary emotion policies.
 - **[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.
-- Vanilla 7B and fixed-emotion 7B baselines.
+- Vanilla 7B (no adapter, no emotion guidance).
 
-Headline result from the paper: **EmoDistill achieves the highest utility across all four domains**, surpassing both vanilla baselines and emotion-selection-only approaches. Full numbers will be cross-linked here when the checkpoint is uploaded.
+**Headline result:** EmoDistill (full) achieves the highest utility across all four domains, surpassing both vanilla and prompt-free baselines, and outperforming the other emotion-aware methods on edge-deployable 7B compute budgets.
 
 ## 📦 Quick start (after checkpoint release)
 
+Loading any variant follows the same pattern — just change the `subfolder` argument:
+
 ```python
 from peft import PeftModel
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 base = "Qwen/Qwen2.5-7B-Instruct"
-adapter = "humanlong/EmoDistill-creditor-7b"
+repo = "humanlong/EmoDistill-7b"
+
+# Pick: ("crad" | "desrd" | "ssad" | "ssd") x ("emodistill" | "promptfree")
+domain  = "crad"
+variant = "emodistill"          # full IQL + SFT + JPO
+# variant = "promptfree"        # LoRA-SFT-only baseline
 
 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))
+model = PeftModel.from_pretrained(model, repo, subfolder=f"{domain}/{variant}/adapter")
 ```
 
-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).
+For the **full pipeline** (IQL emotion selection → LoRA generation → JPO-refined responses), use the helper code in the [EmoDistill GitHub repo](https://github.com/Yunbo-max/EmoDistill):
+
+```python
+from emodistill import EmoDistillAgent
+agent = EmoDistillAgent.from_pretrained("humanlong/EmoDistill-7b", domain="crad")
+reply = agent.respond(conversation_history, opponent_state)
+```
 
 ## ⚠️ Limitations
 
-- Trained for **credit recovery** in English. Generalization to the other three domains (disaster, education, hospital) in the benchmark suite is reported in the paper but not separately released as checkpoints yet.
+- All adapters are trained for **English**. Cross-lingual transfer is not 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.
+- Each adapter is domain-specific — using `crad/emodistill` on a disaster scenario will degrade gracefully but is not the recommended use.
+- The model is designed to be persuasive but ethical — adversarial use to manipulate vulnerable users (debtors, patients, children, disaster survivors) is **out of scope** and explicitly discouraged.
 
 ## 📝 License
 
@@ -120,6 +159,6 @@ Apache 2.0 — matches the base model.
 | [EQ-Negotiator](https://github.com/Yunbo-max/EQ-Negotiator) | NeurIPS 2025 | Personas + HMM + WSLS for SLMs |
 | [EvoEmo](https://github.com/Yunbo-max/EvoEmo) | arXiv preprint | Online evolutionary emotion policies |
 | [EmoMAS](https://github.com/Yunbo-max/EmoMAS) | ACL 2026 (top 9%) | Bayesian multi-agent orchestration + 4 benchmarks |
-| **EmoDistill** *(this repo)* | under review | Offline distillation into a 7B SLM |
+| **EmoDistill** *(this repo)* | under review | Offline distillation: **4 domain models + 4 prompt-free baselines** in a 7B SLM |
 
-🌟 All four in one place: [HF Collection — Emotion-Aware LLM Negotiation](https://huggingface.co/collections/humanlong/emotion-aware-llm-negotiation-6a25d88adcd0b6d41c9d8c75)
+🌟 All five papers + dataset + model in one place: [HF Collection — Emotion-Aware LLM Negotiation](https://huggingface.co/collections/humanlong/emotion-aware-llm-negotiation-6a25d88adcd0b6d41c9d8c75)

crad/emodistill/adapter/README.md

@@ -0,0 +1,9 @@
+# CRAD / emodistill adapter — placeholder
+
+This subfolder will hold the LoRA adapter weights for the **emodistill** variant on the **CRAD** benchmark (Credit Recovery Assessment Dataset (debt recovery)) once training completes.
+
+Files to expect:
+- `adapter_model.safetensors` — LoRA weights
+- `adapter_config.json` — PEFT config
+
+See the top-level [README](../../../README.md) for the full method and loading examples.

crad/emodistill/iql/README.md

@@ -0,0 +1,9 @@
+# CRAD / IQL emotion selector — placeholder
+
+This subfolder will hold the IQL emotion-selector weights for the **CRAD** benchmark (Credit Recovery Assessment Dataset (debt recovery)).
+
+Files to expect:
+- `q_net.pt` — Q-network
+- `v_net.pt` — V-network
+- `policy.pt` — extracted policy
+- `config.json` — emotion vocabulary, IQL hyperparams

crad/promptfree/adapter/README.md

@@ -0,0 +1,9 @@
+# CRAD / promptfree adapter — placeholder
+
+This subfolder will hold the LoRA adapter weights for the **promptfree** variant on the **CRAD** benchmark (Credit Recovery Assessment Dataset (debt recovery)) once training completes.
+
+Files to expect:
+- `adapter_model.safetensors` — LoRA weights
+- `adapter_config.json` — PEFT config
+
+See the top-level [README](../../../README.md) for the full method and loading examples.

desrd/emodistill/adapter/README.md

@@ -0,0 +1,9 @@
+# DESRD / emodistill adapter — placeholder
+
+This subfolder will hold the LoRA adapter weights for the **emodistill** variant on the **DESRD** benchmark (Disaster Emotional Support & Rescue Dataset (emergency)) once training completes.
+
+Files to expect:
+- `adapter_model.safetensors` — LoRA weights
+- `adapter_config.json` — PEFT config
+
+See the top-level [README](../../../README.md) for the full method and loading examples.

desrd/emodistill/iql/README.md

@@ -0,0 +1,9 @@
+# DESRD / IQL emotion selector — placeholder
+
+This subfolder will hold the IQL emotion-selector weights for the **DESRD** benchmark (Disaster Emotional Support & Rescue Dataset (emergency)).
+
+Files to expect:
+- `q_net.pt` — Q-network
+- `v_net.pt` — V-network
+- `policy.pt` — extracted policy
+- `config.json` — emotion vocabulary, IQL hyperparams

desrd/promptfree/adapter/README.md

@@ -0,0 +1,9 @@
+# DESRD / promptfree adapter — placeholder
+
+This subfolder will hold the LoRA adapter weights for the **promptfree** variant on the **DESRD** benchmark (Disaster Emotional Support & Rescue Dataset (emergency)) once training completes.
+
+Files to expect:
+- `adapter_model.safetensors` — LoRA weights
+- `adapter_config.json` — PEFT config
+
+See the top-level [README](../../../README.md) for the full method and loading examples.

ssad/emodistill/adapter/README.md

@@ -0,0 +1,9 @@
+# SSAD / emodistill adapter — placeholder
+
+This subfolder will hold the LoRA adapter weights for the **emodistill** variant on the **SSAD** benchmark (Student Sleep Alerting Dataset (education)) once training completes.
+
+Files to expect:
+- `adapter_model.safetensors` — LoRA weights
+- `adapter_config.json` — PEFT config
+
+See the top-level [README](../../../README.md) for the full method and loading examples.

ssad/emodistill/iql/README.md

@@ -0,0 +1,9 @@
+# SSAD / IQL emotion selector — placeholder
+
+This subfolder will hold the IQL emotion-selector weights for the **SSAD** benchmark (Student Sleep Alerting Dataset (education)).
+
+Files to expect:
+- `q_net.pt` — Q-network
+- `v_net.pt` — V-network
+- `policy.pt` — extracted policy
+- `config.json` — emotion vocabulary, IQL hyperparams

ssad/promptfree/adapter/README.md

@@ -0,0 +1,9 @@
+# SSAD / promptfree adapter — placeholder
+
+This subfolder will hold the LoRA adapter weights for the **promptfree** variant on the **SSAD** benchmark (Student Sleep Alerting Dataset (education)) once training completes.
+
+Files to expect:
+- `adapter_model.safetensors` — LoRA weights
+- `adapter_config.json` — PEFT config
+
+See the top-level [README](../../../README.md) for the full method and loading examples.

ssd/emodistill/adapter/README.md

@@ -0,0 +1,9 @@
+# SSD / emodistill adapter — placeholder
+
+This subfolder will hold the LoRA adapter weights for the **emodistill** variant on the **SSD** benchmark (Surgical Scheduling Dataset (healthcare)) once training completes.
+
+Files to expect:
+- `adapter_model.safetensors` — LoRA weights
+- `adapter_config.json` — PEFT config
+
+See the top-level [README](../../../README.md) for the full method and loading examples.

ssd/emodistill/iql/README.md

@@ -0,0 +1,9 @@
+# SSD / IQL emotion selector — placeholder
+
+This subfolder will hold the IQL emotion-selector weights for the **SSD** benchmark (Surgical Scheduling Dataset (healthcare)).
+
+Files to expect:
+- `q_net.pt` — Q-network
+- `v_net.pt` — V-network
+- `policy.pt` — extracted policy
+- `config.json` — emotion vocabulary, IQL hyperparams

ssd/promptfree/adapter/README.md

@@ -0,0 +1,9 @@
+# SSD / promptfree adapter — placeholder
+
+This subfolder will hold the LoRA adapter weights for the **promptfree** variant on the **SSD** benchmark (Surgical Scheduling Dataset (healthcare)) once training completes.
+
+Files to expect:
+- `adapter_model.safetensors` — LoRA weights
+- `adapter_config.json` — PEFT config
+
+See the top-level [README](../../../README.md) for the full method and loading examples.