reproducibility/README.md

Reproduce the pilot

Use a Linux GPU environment matching environment.txt. This is the exact training source; dependency versions and data hashes are in training.json. Hardware and numerical differences can change the selected checkpoint.

Download the foundation at its pinned revision and this release's synthetic train/validation data. Run from this directory; choose a new output directory.

from huggingface_hub import snapshot_download

snapshot_download(
    "vllm-sr/Vela-1.0-Encoder-307M",
    revision="720ab37904ce15054068d429381bfae7549f00f6",
    local_dir="base",
)
snapshot_download(
    "TwilightTechie/commitment-statements-pilot", repo_type="dataset", local_dir="dataset"
)

python -m sequence_repair.train \
  --method full --fresh-head --trainable-last-layers 2 \
  --base base --base-id vllm-sr/Vela-1.0-Encoder-307M \
  --base-revision 720ab37904ce15054068d429381bfae7549f00f6 \
  --contract ../label-contract.json \
  --train dataset/data/train.jsonl --dev dataset/data/validation.jsonl \
  --output rerun --steps 100 --batch-size 2 --accumulate 2 \
  --max-length 256 --learning-rate 0.00002 --head-learning-rate 0.0001 \
  --eval-every 25 --evaluation-dtype float32 --selection macro-f1 --seed 42

The original run selected step 75 using development macro F1. No reserved test examples are distributed or used by this command. Before running, verify the training and validation file hashes against training.json.