How a 9B‑parameter Qwen3.5 model achieves full‑auto data analysis on a consumer GPU

Model Overview

CoPaw‑Flash‑9B‑DataAnalyst‑LoRA is a LoRA‑fine‑tuned version of Alibaba’s open‑source CoPaw‑Flash‑9B (Qwen3.5‑9B architecture). The LoRA adapter is hosted at huggingface.co/jason1966/CoPaw-Flash-9B-DataAnalyst-LoRA . After fine‑tuning the model can autonomously load CSV/Excel/JSON datasets, perform statistical analysis, generate visualizations, write and execute Python scripts, and produce structured analysis reports without human clicks.

Performance Evaluation

Evaluation used 29 real Kaggle datasets with the Data Analyst framework (max 50 rounds, 128 K context). Results:

Average iteration rounds: 1.2 → 26.0 (≈21.7× increase)

Generated Python files: 0 → >100

Generated charts: 0 → >290

Total token consumption: ~5 K → 18.5 M (≈3700×)

Natural completion rate: 0 % → 89.7 %

Usable outputs: 0/29 (0 %) → 26/29 (90 %)

Human intervention: required at every step → fully autonomous

Demonstration

The agent autonomously analyzes a CSV, writes Python code, executes it, and produces box plots, scatter plots, bar charts, heatmaps, and a final report covering data overview, key findings, dimensional analysis, and conclusions. Sample visualizations (e.g., Toyota used‑car dataset) are shown in the original article.

Deployment Guide

Step 1 – Serve the model with vLLM

export HF_TOKEN=your_huggingface_token

CUDA_VISIBLE_DEVICES=0,1 vllm serve agentscope-ai/CoPaw-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

Key flags: --enable-lora + --lora-modules: load the LoRA adapter (core) --max-lora-rank 64: must match the adapter --reasoning-parser qwen3: expose the model’s reasoning process --enable-auto-tool-choice: automatic tool selection for agent scenarios

Hardware Requirements

Dual‑GPU (bf16, TP=2): ≈11 GB per GPU

Single‑GPU (bf16): ≈22 GB

8‑bit quantization: ≈12 GB

4‑bit quantization: ≈6 GB (consumer‑grade GPU sufficient)

Official test environment: 2 × NVIDIA H200 GPUs with vLLM 0.19.1.

Step 2 – Install the Data Analyst framework

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

Configure .env:

CLAUDE_CODE_USE_OPENAI=1
OPENAI_BASE_URL=http://localhost:8000/v1
OPENAI_API_KEY=unused
OPENAI_MODEL=agent-lora

Step 3 – Run analysis

bun run start

Issue a natural‑language request, e.g.:

Analyze the CSV file in the current directory and find sales trends

The agent loads the data, writes and runs Python code, creates visualizations, and generates a full report automatically.

Model–Framework Relationship

The model acts as the “brain”; the Data Analyst framework provides six tools that translate model intentions into file I/O and code execution. Without the framework the model’s analysis ability cannot be exercised; without LoRA fine‑tuning the original Qwen3.5‑9B model stalls after each tool call, producing no useful output.

Key Takeaways

True autonomy: the agent runs fully automatically, not a step‑by‑step “press‑continue” pseudo‑agent.

9 B parameters are sufficient: consumer‑grade hardware can handle the workload.

All components (model, framework, evaluation data) are released under Apache 2.0.

Empirical results: 90 % success on 29 real datasets demonstrate practical viability.