README.md

---
license: apache-2.0
tags:
- qwen2.5
- lora
- negotiation
---

# Negotiation Models Repository

This repository contains various fine-tuned models for negotiation message reconstruction.

## Available Models

### Qwen3_4B_LORA_1k
- **Location**: `Qwen3_4B_LORA_1k/`
- **Base model**: Qwen/Qwen2.5-4B-Instruct
- **Method**: LoRA (Low-Rank Adaptation)
- **Training samples**: ~1k negotiation examples
- **Hardware**: 2x A100 GPUs
- **Size**: ~505MB (LoRA adapter only)

### Qwen3_4B_Full_1k
- **Location**: `Qwen3_4B_Full_1k/`
- **Base model**: Qwen/Qwen2.5-4B-Instruct
- **Method**: Full fine-tuning (all parameters)
- **Training samples**: ~1k negotiation examples
- **Hardware**: 2x A100 GPUs with DeepSpeed ZeRO-3
- **Size**: ~7.6GB (full model weights)

## Usage

### Using LoRA model
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-4B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-4B-Instruct")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "JLiangHe/negotiation_clone", subfolder="Qwen3_4B_LORA_1k")
```

### Using Full fine-tuned model
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load full fine-tuned model directly
model = AutoModelForCausalLM.from_pretrained("JLiangHe/negotiation_clone", subfolder="Qwen3_4B_Full_1k")
tokenizer = AutoTokenizer.from_pretrained("JLiangHe/negotiation_clone", subfolder="Qwen3_4B_Full_1k")
```

## Model Comparison

| Model | Method | Size | Memory | Performance |
|-------|--------|------|--------|-------------|
| Qwen3_4B_LORA_1k | LoRA | 505MB | Low (base + adapter) | Good for inference efficiency |
| Qwen3_4B_Full_1k | Full FT | 7.6GB | High (full model) | May have better task adaptation |