| #!/usr/bin/env bash |
| set -euo pipefail |
|
|
| PROJECT_DIR="${PROJECT_DIR:-/workspace/KAWK50M}" |
| RECOVERY_CONFIG="${RECOVERY_CONFIG:-configs/pretrain_fixed_recovery_100m_5090.yaml}" |
| SCRATCH_CONFIG="${SCRATCH_CONFIG:-configs/pretrain_fixed_scratch_100m_5090.yaml}" |
| BROKEN_CHECKPOINT="${BROKEN_CHECKPOINT:-runs/kawk-1.5-50m-v2-cpt-3b-5090/checkpoint-00024415}" |
| RECOVERY_RUN="${RECOVERY_RUN:-runs/kawk50m-fixed-recovery-100m-5090}" |
| SCRATCH_RUN="${SCRATCH_RUN:-runs/kawk50m-fixed-scratch-100m-5090}" |
| REPORT_DIR="${REPORT_DIR:-reports/fixed-objective-100m-ab}" |
| VALIDATION_BIN="${VALIDATION_BIN:-data/tokenized-v2b-100m/validation.bin}" |
| CLEAN_DIR="${CLEAN_DIR:-data/clean-v2-full}" |
|
|
| cd "${PROJECT_DIR}" |
| source "${PROJECT_DIR}/.venv/bin/activate" |
|
|
| for path in "${RECOVERY_RUN}" "${SCRATCH_RUN}" "${REPORT_DIR}"; do |
| if [[ -e "${path}" ]]; then |
| echo "Refusing to overwrite existing A/B artifact: ${path}" >&2 |
| exit 1 |
| fi |
| done |
| if [[ ! -f "${BROKEN_CHECKPOINT}/model/model.safetensors" ]]; then |
| echo "Recovery source checkpoint is missing: ${BROKEN_CHECKPOINT}" >&2 |
| exit 1 |
| fi |
|
|
| python -m scripts.preflight --config "${RECOVERY_CONFIG}" |
| python -m scripts.preflight --config "${SCRATCH_CONFIG}" |
| mkdir -p "${REPORT_DIR}" |
|
|
| python -u -m scripts.pretrain \ |
| --config "${RECOVERY_CONFIG}" \ |
| --init-from "${BROKEN_CHECKPOINT}" \ |
| 2>&1 | tee "${REPORT_DIR}/recovery-train.log" |
|
|
| python -u -m scripts.pretrain \ |
| --config "${SCRATCH_CONFIG}" \ |
| 2>&1 | tee "${REPORT_DIR}/scratch-train.log" |
|
|
| recovery_name="$(cat "${RECOVERY_RUN}/latest_checkpoint.txt")" |
| scratch_name="$(cat "${SCRATCH_RUN}/latest_checkpoint.txt")" |
| recovery_checkpoint="${RECOVERY_RUN}/${recovery_name}" |
| scratch_checkpoint="${SCRATCH_RUN}/${scratch_name}" |
|
|
| common_evaluation=( |
| --validation-bin "${VALIDATION_BIN}" |
| --sequence-length 1024 |
| --batch-size 32 |
| --eval-batches 20 |
| --max-new-tokens 100 |
| --temperature 0.7 |
| --top-k 50 |
| --top-p 0.9 |
| --repetition-penalty 1.15 |
| --seed 42 |
| ) |
|
|
| python -m scripts.evaluate \ |
| --checkpoint "${BROKEN_CHECKPOINT}" \ |
| "${common_evaluation[@]}" \ |
| > "${REPORT_DIR}/broken-final-evaluation.json" |
|
|
| python -m scripts.evaluate \ |
| --checkpoint "${recovery_checkpoint}" \ |
| "${common_evaluation[@]}" \ |
| > "${REPORT_DIR}/recovery-evaluation.json" |
|
|
| python -m scripts.evaluate \ |
| --checkpoint "${scratch_checkpoint}" \ |
| "${common_evaluation[@]}" \ |
| > "${REPORT_DIR}/scratch-evaluation.json" |
|
|
| python -m scripts.evaluate_text \ |
| --checkpoint "${recovery_checkpoint}" \ |
| --input "${CLEAN_DIR}/validation-*.jsonl" \ |
| --context-length 1024 \ |
| --max-documents 200 \ |
| > "${REPORT_DIR}/recovery-byte-evaluation.json" |
|
|
| python -m scripts.evaluate_text \ |
| --checkpoint "${scratch_checkpoint}" \ |
| --input "${CLEAN_DIR}/validation-*.jsonl" \ |
| --context-length 1024 \ |
| --max-documents 200 \ |
| > "${REPORT_DIR}/scratch-byte-evaluation.json" |
|
|
| printf '%s\n' "${recovery_checkpoint}" > "${REPORT_DIR}/recovery-checkpoint.txt" |
| printf '%s\n' "${scratch_checkpoint}" > "${REPORT_DIR}/scratch-checkpoint.txt" |
| echo "Correct-objective 100M recovery/scratch A/B training and evaluation complete." |
|
|
