Reproducibility¶

The paper-release v4 dataset is pinned, checksummed, and verified byte-identical against the trained-on file. To regenerate it from scratch you need:

The pinned generation commit: 898814d. Later releases include changes that came after this commit — most notably the noun_case arc gate. Generating against any release tag will not produce bit-identical output.
The exact source corpora: documented in data/V4_DATA_PROVENANCE.md.
Seed=42 in both build_v4_sources.py and generate_sft.py.

Verifying you have the right files¶

uv run python scripts/verify_v4.py

This checks SHA256 against data/v4_checksums.txt. Output should be:

8/8 OK, 0 mismatch, 0 missing

Regenerating from scratch¶

git checkout 898814d

# 1. Mine scarce sentences
uv run python scripts/mine_scarce_sents.py …

# 2. Extract clean rublimp pool
uv run python scripts/extract_rublimp_pool.py …

# 3. Build source mix (150K, 60-40 pool/news split)
uv run python scripts/build_v4_sources.py \
    --output data/mixed_sources_v4.txt \
    --total 150000 --seed 42

# 4. Generate SFT
uv run python scripts/generate_sft.py \
    -i data/mixed_sources_v4.txt \
    -o data/qwen_sft_v4.jsonl \
    -n 50000 --seed 42 --depparse \
    --max-input 150000 --batch-size 128 \
    --balance-directions

Full step-by-step (with corpus paths and benchmark exclusions) is in V4_DATA_PROVENANCE.md.

What the v4 dataset is¶

39,209 SFT examples across the synterr handler set
Source mix: 54,823 scarce-form-mined sentences + 57,106 RuBLiMP pool + 38,071 Taiga news
No RuBLiMP benchmark contamination (excluded at build time)
Direction-balanced for split / merge / insert / delete handlers

Citing¶

When citing the dataset specifically (vs. the synterr tool), reference the paper and the pinned generation commit 898814d.