# Mol-LLM data

Use the accompanying submission code at commit ae8a2acbb9a521f919f9b50876f83cdfbf43c713.

All training upsampling repetitions are removed: only `molpo_dup_idx == 0` remains. There are 2,818,007 training rows and 152,482 evaluation rows. Identical records already present in the original datasets are preserved. This is an unupsampled release; the supplied trained checkpoints retain their original training history.

`train/` and `test/` contain prepared Parquet partitions. `sources/train/` and `selections/` reconstruct the training partition through `python -m data.builders.hard_negative --source /path/to/bundle/sources --pairs /path/to/bundle/selections --output /path/to/new/train --workers 8`. Selection indices refer to the unchanged canonical TRAIN partition. Each retained instance keeps all three reject slots.

Run `sha256sum -c SHA256SUMS` before loading. `validation.json` describes verification coverage. There is no separate validation partition.