---
license: apache-2.0
base_model: Qwen/Qwen2.5-0.5B-Instruct
library_name: transformers
tags:
- dataforge
- data-quality
- supervised-fine-tuning
- qlora
- kaggle
datasets:
- Praneshrajan15/dataforge-sft-trajectories
metrics:
- f1
model-index:
- name: DataForge-0.5B-SFT
  results:
  - task:
      type: text-generation
      name: DataForge repair planning
    dataset:
      name: DataForge SFT trajectories
      type: Praneshrajan15/dataforge-sft-trajectories
    metrics:
    - type: macro_f1
      name: Held-out macro F1
      value: null
---

# DataForge-0.5B-SFT

DataForge-0.5B-SFT is a supervised-fine-tuned warmup checkpoint for tabular
data-quality repair experiments. The current training path uses chunk-level
DataForge expert trajectories whose exact repairs are derived from audited
dirty/clean CSV diffs. The earlier `v0-smoke` release only proved the
Kaggle-to-Hugging-Face pipeline and should not be read as a performance claim.

## Intended Use

- Research on tabular data-quality agents and repair planning.
- Offline evaluation on DataForge-Bench-style Hospital, Flights, and Beers
  tasks.
- Warm-starting later DataForge RL experiments.

This checkpoint is not intended for autonomous production data modification,
medical decision support, regulated data governance, or unsupervised repair of
private datasets.

## Training Data

- Dataset repo: `Praneshrajan15/dataforge-sft-trajectories`.
- Dataset repo SHA used for this run: `94e2dd556d4f1260c5123d93ca6bf4f9da9b160a`.
- Training examples: `1958` chunk-level `expert_v4.jsonl`
  records.
- Data sources: Raha benchmark Hospital, Flights, and Beers datasets via the
  BigDaMa/raha repository.
- Primary label source: `oracle_from_clean_diff` dirty/clean CSV diffs.
- Legacy teacher lineage: Groq-hosted `clean-diff-v1` ReAct smoke records may
  remain for auditability, but exact repairs are not teacher-discovered labels.
- Flights schedule and actual-time repairs are supervised from dirty/clean
  labels; they are not inferred from incomplete prompt context.
- Split safety: held-out rows are reserved before chunking and excluded from SFT
  target rows, context rows, normalization candidates, fixes, and messages.
- Hard negatives: clean train chunks are retained as `finish` examples with
  empty repairs so the model is penalized for unnecessary edits.

The trajectory JSONL includes state, tool calls, diagnosis text, proposed fixes,
teacher/oracle metadata, benchmark metrics, split metadata, and source
provenance for auditability.

## Training Procedure

- Base model: `Qwen/Qwen2.5-0.5B-Instruct`.
- Method: 4-bit QLoRA warmup, then LoRA merge into fp16 merged weights.
- Compute target: Kaggle or Hugging Face remote GPU only; no laptop model
  training or full evaluation.
- Kaggle hours used: `1.147`.
- Epochs: 2.
- Batch size: 1 per device with gradient accumulation of 16.
- Learning rate: 2e-5.

## Evaluation

Evaluation is reported on held-out DataForge-Bench-style tasks sampled after the
training trajectory seeds. The release status generated by the notebook is
`quality_improved_verified`. Only `quality_improved_verified` should be treated as a
quality milestone. `diagnostic_complete_no_gain` means the run is authentic and
published, but not promoted.

| Model | Held-out macro F1 |
| --- | ---: |
| `Qwen/Qwen2.5-0.5B-Instruct` | `0.0` |
| `DataForge-0.5B-SFT` | `0.0077` |

Release gates:

- Parse success: `1.0`.
- Schema-case errors: `0`.
- Quality milestone: `True`.

These numbers are produced by the publishing notebook and should not be edited
manually. Re-run the notebook to regenerate them. Detailed per-dataset metrics
are stored in `training_metrics.json` under `base_eval` and `sft_eval`.
Bounded per-task failure evidence is stored in `eval_diagnostics.json`.

## Limitations

- The checkpoint is a Week 9 warmup model, not the final DataForge model family.
- It has only seen small chunk-level ReAct traces and may fail on larger schemas,
  unseen domains, adversarial dirty values, or tasks requiring multi-step
  database access.
- Legacy teacher traces can contain teacher errors; the primary current labels
  come from exact dirty/clean diffs.
- The model should be used behind DataForge's safety, verifier, and transaction
  layers before any real data changes.

## License

Weights are published as `apache-2.0` after verifying the base model
metadata for `Qwen/Qwen2.5-0.5B-Instruct`. Users must also comply with the source dataset
licenses/terms and the teacher model terms that governed trajectory generation.