Praneshrajan15/DataForge-0.5B-SFT

DataForge-0.5B-SFT: A Warmup Model for Tabular Data Quality Repair

DataForge-0.5B-SFT is a specialized 0.5 billion parameter model developed by Praneshrajan15, fine-tuned from Qwen/Qwen2.5-0.5B-Instruct. Its primary purpose is to serve as a supervised-fine-tuned warmup checkpoint for experiments in tabular data-quality repair.

Key Capabilities & Training

• Tabular Data Repair: The model is trained on chunk-level DataForge expert trajectories, learning exact repairs derived from audited dirty/clean CSV diffs across datasets like Hospital, Flights, and Beers.
• Warm-starting RL Experiments: It is intended to warm-start later DataForge Reinforcement Learning (RL) experiments, focusing on data-quality agents and repair planning.
• Supervised Fine-Tuning: Trained using 4-bit QLoRA warmup followed by LoRA merge into fp16 weights, on 1958 chunk-level expert_v4.jsonl records from Praneshrajan15/dataforge-sft-trajectories.
• Evaluation: Achieved a held-out macro F1 score of 0.0077 on DataForge-Bench-style tasks, improving from 0.0 for the base model.

Intended Use Cases

• Research: Ideal for research into tabular data-quality agents and repair planning.
• Offline Evaluation: Suitable for offline evaluation on DataForge-Bench-style tasks.
• Warm-starting: Designed to warm-start subsequent DataForge RL experiments.

Limitations

As a Week 9 warmup model, DataForge-0.5B-SFT has limitations. It has only processed small chunk-level ReAct traces and may not perform well on larger schemas, unseen domains, adversarial dirty values, or tasks requiring multi-step database access. It is not intended for autonomous production data modification or unsupervised repair of private datasets.