Blame for README.md

---
language: en
license: apache-2.0
task_categories:
- text-generation
tags:
- code
- pandas
- data-analysis
size_category: small
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: eval
    path: eval.jsonl
---

# analyst-data

Training corpus for analyst-v1: a small pandas-codegen model.

- 613 train / 165 eval-pool chat-format pairs (eval pool = pairs with constructible test DataFrames, held out of training)
- Sources: DS-1000 seed (real Kaggle-style data problems) + synthetic pairs generated from live Kaggle datasets (every pair execution-verified against the real DataFrame)

Format: {"system": ..., "user": "Schema:... Question:...", "assistant": "<pandas code>"}