---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
tags:
- pandas
- code-generation
- data-analysis
- lora
---

# analyst-v1

A LoRA adapter for **Qwen2.5-Coder-1.5B-Instruct** that writes short pandas code answering questions about a DataFrame `df` already loaded in scope.

Trained on 613 (schema, question, code) pairs. Every training target was **executed against the real DataFrame before being kept**: 546 pairs derived from DS-1000, 67 generated against real public Kaggle CSVs. See the companion paper: *Every Answer Must Run: Training Small Data-Analyst Models Under Execution Verification*.

## Measured (150 held-out exec-scored pairs)

| arm | exec_rate | exact_match |
|---|---|---|
| Qwen2.5-Coder-1.5B-Instruct | 0.533 | 0.293 |
| + this adapter | 0.620 | 0.253 |

Execution rises 8.7 points; exact match against bare-script references falls 4.0 points (the adapter adopts a wrapper-function answer style; see paper section 6 for the mechanism).

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-1.5B-Instruct", torch_dtype="bfloat16").cuda()
model = PeftModel.from_pretrained(base, "em908/analyst-v1")
model = model.merge_and_unload().eval()
```

## Training data

`em908/analyst-data` (same pair counts, same execution filter).

## License

Apache-2.0 (matches the base model).