Reproducibility¶

The paper-release v4 dataset is pinned, checksummed, and verified byte-identical against the trained-on file. Reproducibility is archival, not regenerative: the checksummed files are the authoritative objects, and the pipeline below does not re-create them bit for bit.

The pinned generation commit: tag v4-data (commit 2fd4d78; pinned as 898814d before the May 2026 history rewrite). Later releases change handler behavior — most notably the noun_case arc gate — so generating against any release tag produces different output by design.
The source corpora: documented in data/V4_DATA_PROVENANCE.md. The shipped source mix mixed_sources_v4.txt was not produced by the committed build_v4_sources.py and cannot be regenerated from any committed script (see the provenance caveat there); it is archived and checksummed as-is.
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

Re-running the intended pipeline¶

git checkout v4-data

# 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
Intended source mix: 54,823 scarce-form-mined sentences + 57,106 RuBLiMP pool + 38,071 Taiga news, RuBLiMP benchmark items excluded. The shipped mix differs: 154,806 non-blank lines, only 107,265 unique, so the SFT data likely repeats some source sentences (a documented v4 limitation)
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 (tag v4-data).