Arki05's picture
docs: top-level README + mixes/ (catalog, per-mix configs, mix_dataset.py builder)
d5fd5b6 verified
|
Raw
History Blame Contribute Delete
2.87 kB

Training mixes

Each mix is a weighted, shuffled, materialized set of token shards (.bin/.idx/.desc + manifest.json, per molvae.data.format), built from ../raw/ by mix_dataset.py. The shards live on S3 (s3://mechanophore/molvae/tokens/<name>/) and are mirrored under each mix's data/. The *.config.json here is the exact, re-runnable spec.

Builder

mix_dataset.py is torch-free (numpy + pyarrow + boto3 + huggingface_hub). Disk-backed & bounded-RAM: Phase A streams each source once and fraction-samples it to disk (vectorized per parquet batch — ~7M rows/s; parquets are downloaded, not fsspec-streamed); Phase B external scatter-shuffles the per-source temp shards into --out-shards output buckets. Per-source knobs: weight, min_len/max_len (length filter, e.g. long-molecule streams — keep-prob is calibrated to the post-filter pool), and for bin sources a size-stratified shard selection with per_shard_cap (or default largest-first). Every build dedups against the held-out mechanophore test set (--dedup-holdout).

# rebuild a mix (on the CPU node; source S3 creds first)
python mix_dataset.py --config mix-v2-1B.config.json \
    --dedup-holdout mechano_test.jsonl --out /tmp/mix \
    --s3 s3://mechanophore/molvae/tokens/mix-v2-1B --out-shards 512 --shard-size 100000

Design principle

Sample by weights, not raw counts — the mechanophore target is a strained core at ~27 heavy atoms with esters/nitriles, which no single source spans. GDB gives strained ring grammar (but small); mcule/eMolecules give the long molecules (but drug-like); ZINC gives a medium size anchor. The mix's job is to span the composition. (Long+strained together still needs Phase-2 sources — COCONUT / SAVI — not yet added.)

mix-v2-1B — the strained-rich pretrain mix (current)

~60% GDB (strain) · ~40% size/length · ~9–13% of mols ≥55 tokens (long-decode).

weight source role
35% gdb13 strain-volume workhorse (≤13 HA rings)
8% GDB4c3D curated strained
8% GDB-13s ring diversity
4% gdb11 small exhaustive
5% GDBMedChem drug-like ring
12% ZINC (largest-first) medium size anchor
10% mcule long-ish (median 46)
5% mcule-long (min_len 55) long-decode leg (all ≥55, to 154 tok)
13% eMolecules ordersc long + diverse (to 168 tok)

mix-v2-40M is the same composition at 40M (pilot). mix-phase1-40M is the earlier v1 (ZINC 40% / GDB4c3D 35% / GDBMedChem 15% / ChEMBL 10%). See each *.config.json for exact specs.

Loading (training side, needs torch)

from molvae.data import MolvaeDataset, make_loader
ds = MolvaeDataset("<local>/mix-v2-1B", "train")      # mmap token shards
loader = make_loader(ds, pad_id=3, token_budget=12288)