ATOM (atom-ai): code snapshot (100fae0 + batched-train), README with measured numbers, curated checkpoints (part 4)
b9301f1 verified Download scripts/train_decfix.sh from thefinalboss/atom-ai: direct link, hf CLI and curl.
- Browser
- Download file 1.95 kB
-
https://huggingface.co/thefinalboss/atom-ai/resolve/main/scripts/train_decfix.sh
- Command line
-
hf download hf://thefinalboss/atom-ai/scripts/train_decfix.sh
-
curl -L -o train_decfix.sh https://huggingface.co/thefinalboss/atom-ai/resolve/main/scripts/train_decfix.sh
1.95 kB
| # STEP 1 of docs/DECODE_FIX_AND_BPTT.md: short fine-tune from fieldsat 4996k with the | |
| # byte-tick decode-label fix active and --efference-every 0 (removes the copy rule the | |
| # off-by-one efference taught). Exact fieldsat recipe otherwise. New dir, new suffix. | |
| # env START=4996000 END_TARGET=5012000 CHUNK=8000 bash scripts/train_decfix.sh | |
| set -eu | |
| cd /workspace/exports/atom-ai | |
| START=${START:-4996000}; END_TARGET=${END_TARGET:-5012000}; CHUNK=${CHUNK:-8000} | |
| LR=${LR:-1.5e-5}; SLR=${SLR:-3.6e-5}; BASE=496000 | |
| OUT=checkpoints/byte_tick_decfix; SUF=d32_ms1M_lrd_b4828_fieldsat_decfix | |
| LOGD=logs/decfix; mkdir -p "$OUT" "$LOGD" | |
| export PYTHONPATH=. OMP_NUM_THREADS=${THREADS:-1} MKL_NUM_THREADS=${THREADS:-1} | |
| CUR=$START | |
| RESUME=checkpoints/byte_tick_fieldsat/atom_native_step_${START}_d32_ms1M_lrd_b4828_fieldsat.pt | |
| while [ "$CUR" -lt "$END_TARGET" ]; do | |
| END=$((CUR + CHUNK)); NAME="atom_native_step_${END}_${SUF}.pt" | |
| if [ ! -f "$OUT/$NAME" ]; then | |
| echo "$(date -Is) chunk $CUR->$END resume=$RESUME" | |
| .venv/bin/python -u tools/run_atom_native.py \ | |
| --stream --data-glob 'data/corpus_fr_medium*.txt' --loop-shards --chunk-bytes 65536 \ | |
| --stream-skip-packets $((CUR - BASE)) --resume "$RESUME" --output-dir "$OUT" \ | |
| --checkpoint-name "$NAME" --steps "$CHUNK" \ | |
| --d-model 32 --n-modes 32 --n-atoms-max 64 --max-span-bytes 1 \ | |
| --field-obligatory-hard --no-enable-merge --no-payload-copy --last-atom-readout \ | |
| --efference-every 0 --learning-rate "$LR" --surface-learning-rate "$SLR" --seed 20260913 \ | |
| --atom-flush-every 64 --episode-length 8000 --episode-reset --slow-every 1 \ | |
| --field-loss-weight 0.05 --field-contrast-weight 0 --allow-parallel-train \ | |
| --last-atom-dropout 0.5 --field-leak 0.3 > "$LOGD/train_${CUR}_${END}_decfix.log" 2>&1 | |
| cp "$OUT/run_metrics.json" "$OUT/run_metrics_${END}_${SUF}.json" | |
| fi | |
| RESUME="$OUT/$NAME"; CUR=$END | |
| done | |
| echo "$(date -Is) done" | |