---
license: apache-2.0
language:
- en
- zh
base_model:
- agentscope-ai/CoPaw-Flash-9B
tags:
- agent
- tool-use
- lora
- vllm
- qwen3.5
pipeline_tag: image-text-to-text
library_name: peft
---

# CoPaw-Flash-9B-DataAnalyst-LoRA

[![Demo Video](https://img.shields.io/badge/Demo_Video-YouTube-red?style=for-the-badge&logo=youtube)](https://youtu.be/A0WBHV-DzIE?si=_sHpjOUuhzjmN43p)
[![Showcase](https://img.shields.io/badge/Showcase-29_Datasets-blue?style=for-the-badge&logo=google-chrome&logoColor=white)](https://dataanalyst.locoremind.com/)
[![Framework](https://img.shields.io/badge/Framework-Data_Analyst-green?style=for-the-badge&logo=github)](https://github.com/IIIIQIIII/data-analyst)
[![vLLM](https://img.shields.io/badge/Deployment-vLLM-orange?style=for-the-badge)](https://docs.vllm.ai/)

**Agentic Data Analyst that autonomously explores, analyzes, and visualizes your datasets.**

<p align="center">
  <img src="dataanalyst-demo.gif" alt="Data Analyst Demo" width="800">
</p>

## What It Does

This model functions as an **autonomous data analyst**:
- 📂 Loads and explores datasets (CSV, Excel, JSON)
- 🔍 Performs statistical analysis and data profiling
- 📊 Creates visualizations (matplotlib, seaborn, plotly)
- 🐍 Writes and executes Python analysis scripts
- 📝 Generates summary reports and insights
- 🔄 Iterates through multi-step analysis workflows
- 🎯 Completes 90% of tasks autonomously (no human intervention)

## Model Details

| Property | Value |
|----------|-------|
| Base Model | agentscope-ai/CoPaw-Flash-9B (Qwen3.5-9B architecture) |
| Task Type | Data Analysis Agent |
| LoRA Rank | 64 |
| LoRA Alpha | 128 |
| Precision | bfloat16 |
| PEFT Version | 0.18.1 |

## Performance Benchmark

Tested on **29 real Kaggle datasets** with [Data Analyst](https://github.com/IIIIQIIII/data-analyst) framework (max_turns=50, context=128K):

| Metric | Qwen3.5-9B Base | DataAnalyst-LoRA | Improvement |
|--------|-----------------|------------------|-------------|
| **Avg iterations** | 1.2 | 26.0 | **21.7x** |
| **Python files** | 0 | 100+ | **∞** |
| **Charts generated** | 0 | 290+ | **∞** |
| **Total tokens** | ~5K | 18.5M | **3700x** |
| **Natural completion rate*** | 0% | 89.7% | **+89.7pp** |
| **Hit turn limit** | N/A | 10.3% | - |
| **Usable output** | 0/29 (0%) | 26/29 (90%) | **+90pp** |
| **User intervention** | Required every step | Autonomous | **Autonomous** |

*Natural completion = Model autonomously outputs final summary report within 50 turns

<p align="center">
  <img src="benchmark-chart.png" alt="Performance Benchmark: Base Model vs DataAnalyst-LoRA" width="800">
</p>

### Key Findings

**Base Model (Qwen3.5-9B):**
- ❌ Understands tool call format but cannot execute autonomously
- ❌ Stops after 1-2 iterations
- ❌ Requires continuous user "continue" prompts
- ❌ Produces zero analysis output
- ❌ Not usable for real data analysis tasks

**CoPaw-Flash-9B-DataAnalyst-LoRA:**
- ✅ Fully autonomous execution (26 iterations average)
- ✅ Generates complete analysis pipelines
- ✅ Creates visualizations and reports
- ✅ 90% success rate on real-world datasets
- ✅ Production-ready for data analysis workflows

**Conclusion:** LoRA training is **essential, not optional**. Base model lacks autonomous data analyst capabilities despite understanding the tool calling format. This LoRA transforms the base model into a production-ready AI data analyst that can handle real-world datasets independently.

## Quick Start

### Step 1: Deploy with vLLM

```bash
export HF_TOKEN=your_huggingface_token

CUDA_VISIBLE_DEVICES=0,1 vllm serve agentscope-ai/QwenPaw-Flash-9B \
  --enable-lora \
  --lora-modules agent-lora=jason1966/CoPaw-Flash-9B-DataAnalyst-LoRA \
  --max-lora-rank 64 \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.85 \
  --max-model-len 131072 \
  --gdn-prefill-backend triton \
  --trust-remote-code \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --port 8000
```

### Step 2: Setup Data Analyst Framework

```bash
git clone https://github.com/IIIIQIIII/data-analyst.git
cd data-analyst
bun install
```

Configure `.env`:
```bash
CLAUDE_CODE_USE_OPENAI=1
OPENAI_BASE_URL=http://localhost:8000/v1
OPENAI_API_KEY=unused
OPENAI_MODEL=agent-lora
```

### Step 3: Start Analyzing

```bash
bun run start
```

Then input your analysis task:
```
Analyze sales_2024.csv and identify trends
```

The model will autonomously load data, perform analysis, create visualizations, and generate reports—all without requiring manual "continue" prompts.

## vLLM Parameters

| Parameter | Description |
|-----------|-------------|
| `--enable-lora` | Enable LoRA adapter support |
| `--lora-modules agent-lora=...` | Load DataAnalyst-LoRA adapter |
| `--max-lora-rank 64` | LoRA rank (must match adapter) |
| `--reasoning-parser qwen3` | Enable reasoning process visibility |
| `--enable-auto-tool-choice` | Automatic tool selection |
| `--tool-call-parser qwen3_xml` | Parse XML-format tool calls |
| `--gdn-prefill-backend triton` | Optimize prefill with Triton |

## Hardware Requirements

| Configuration | VRAM Required |
|--------------|---------------|
| Dual GPU (bf16, TP=2) | ~11GB per GPU |
| Single GPU (bf16) | ~22GB |
| 8-bit quantized | ~12GB |
| 4-bit quantized | ~6GB |

**Tested:** 2x NVIDIA H200, vLLM 0.19.1, CUDA 13.0, Python 3.12

## Troubleshooting

| Issue | Solution |
|-------|----------|
| FlashInfer errors | Add `--gdn-prefill-backend triton` |
| Out of memory | Reduce `--max-model-len` or `--gpu-memory-utilization` |
| Connection refused | Check `netstat -tlnp \| grep 8000` |

## Acknowledgments

- [CoPaw-Flash-9B](https://huggingface.co/agentscope-ai/QwenPaw-Flash-9B) — Base model by AgentScope AI
- [Brev.dev](https://brev.nvidia.com/) — GPU cloud infrastructure by NVIDIA
- [LocoreMind](https://locoremind.com/) — Research and development

## License

Apache 2.0