mpl_eff_lr

# MPL Effective LR Focused Training Curves This dataset contains the focused runnable subset for the `mpl_eff_lr` code release. Included conditions: - `1b7 / muonh / fp8 / bs512` - `1b7 / muonh / mxfp8 / bs1536` - peak LR values `0.005`, `0.006`, `0.007` - `lr_mult = 3` - target steps `10000`, `20000`, `40000` - schedule `linear` Rows in manifest: 18 Each run directory contains: ```text meta.json curve.csv ``` `curve.csv` columns: ```text step,learning_rate,effective_lr_matrix,training_loss,validation_loss,consumed_samples,timestamp,source_run_key ``` Use with the GitHub repo: ```bash hf download thu-pacman/mpl_eff_lr --repo-type dataset --local-dir data --include "runs/**" "manifest.csv" "README.md" python -m scripts.pipeline.run --config configs/pipelines/1b7_fp8_bs512_cross_peak.yaml --dry-run ```