Data & Training

NovaEdit is trained to emit patch DSL from code regions + diagnostics. This repo includes scaffolding you can extend with real data and checkpoints.

Data layout

• data/python/raw — raw corpora
• data/python/diffs — mined git diffs
• data/python/synthetic_bugfixes — generated bugfix samples
• data/javascript/* — future languages

Prep scripts

bash scripts/prepare_python_data.sh
# mines small diffs and synthetic samples into data/python/...

• scripts/mine_git_diffs.py — collect recent Python diffs from a repo.
• scripts/generate_synthetic_bugs.py — inject varied bugs (missing imports, typos, off-by-one, comparators, missing returns) and save JSONL.
• scripts/build_edit_dataset.py — merge multiple JSONL sources into one.
• scripts/train_tokenizer.py --input-glob 'data/python/raw/**/*.py' --output model/tokenizer.json — train a BPE tokenizer.
• Use --validate on generate_synthetic_bugs.py to skip samples that fail AST parsing.
• scripts/split_dataset.py --input data/python/processed/edits.jsonl --train-out ... — split a merged dataset into train/val/test.

Training stubs

• trainer/pretrain.py — tiny character LM to smoke-test pipelines.
• trainer/sft_edit.py — SFT scaffold using the heuristic model as pseudo-labels.
• Config examples live in model/config/*.yaml (small/base).

Evaluation

• eval/run_eval_bugfix.py --data <jsonl> — measures diagnostic count reduction.
• eval/run_eval_regression.py — prints patches for a small regression suite.

Hugging Face

• Push weights/config with scripts/push_to_hub.py --repo <org/model> --path weights/novaedit-small.
• For Spaces, run the FastAPI app (novaedit.server.main:app) via the provided Dockerfile.

Samples

• data/python/processed/sample_edits.jsonl — tiny, ready-to-use edit dataset for smoke tests.
• model/tokenizer-sample.json — toy tokenizer generated from examples/ for pipeline checks (train your own for real runs).