pooper-b6 (407M, 128 kb context)

A ~407M-parameter model that predicts 1,885 ENCODE signal tracks (ATAC-seq, DNase-seq, histone ChIP-seq, RNA-seq; GRCh38) at 1 bp resolution from a 131,072 bp DNA window. It uses an AlphaGenome-style backbone trained from random initialisation (no AlphaGenome pretrained weights) with a per-track linear head.

Training supervision comes from the Zipoop teacher shards: one over-fitted small "teacher" network per 32,768 bp window, each compressing that window's 1,885 tracks.

Model details

Trainable parameters	406,817,601 (the checkpoint also holds unused frozen AlphaGenome modules; 453M tensor parameters in total)
Input	131,072 bp (4 x 32,768 bp windows)
Output	1,885 tracks, 1 bp resolution
Loss	Pearson loss, plus a small (weight 0.1) scaled-space MSE term
Optimiser	lr peak 5e-4, 500 warm-up steps, cosine decay to 10%, bf16, seed 0
Schedule	20,000 steps, batch of 4 windows per step, gradient checkpointing
Training split	AlphaGenome fold0 split (63,946 train / 10,822 valid / 10,345 test windows), used only to keep evaluation comparable

Training ran in two stages: steps 0-3,000 as the "B6" control experiment, then steps 3,000-20,000 resumed from that checkpoint with the cosine schedule re-laid for 20,000 total steps. Both logs are in repro/config/.

Results

Validation (fold0, per-window Pearson, from the training log, final step 20,000):

Assay	Pearson	log1p 128 bp Pearson
ATAC	0.674	0.611
DNase	0.653	0.635
ChIP (histone)	0.447	0.452
RNA-seq	0.290	0.558
mean / balanced	0.416 / 0.516	- / 0.564

A sibling model with the same recipe (b8) was additionally evaluated with the AlphaGenome-style per-track pooled Pearson on 400 test windows; see the b8 repository.

Files

• pooper_genomeb6_407M_128K.pt: state_dict, config, args, track stats and step (1,814,125,434 bytes, SHA256 9f5e800e1151992b644d5092843b4ae2284bb0b875da0dc5ab6edb7ce48c64f0). No optimiser state.
• repro/: training and evaluation code (code.tgz), the exact queue scripts, start events, validation curve, full training log, fold0 split, window index checksums and library versions. See repro/README_复现.md (in Chinese) for the reproduction steps. The 120 GB teacher shards are not included.

Limitations

• Metrics above are validation-set numbers from the training log. An official-style fold1 test evaluation for b6 itself is not included here.
• RNA-seq and histone ChIP are clearly weaker than ATAC and DNase.
• This is a research prototype, not a validated tool for clinical or variant-effect use.
• The model was trained against Zipoop teacher shards; this card does not claim that scores reflect agreement with raw ENCODE signal.
• Results come from a single seed.