linear-intestine

linear trained or fitted separately on intestine, using 19331 canonical input positions and the frozen split 5,973 train / 6,596 validation / 5,541 test.

Scratch training: seed 0, effective batch 128, 20 complete epochs, all layers trainable; no pretrained weights. The epoch20 endpoint and best-validation-CE endpoint are separate. checkpoints/ contains original native .pt files with optimizer/RNG and scientific endpoints; inference/ contains tensor-only safetensors exports checked bitwise against the original states. Linear endpoints that are byte-identical are represented by aliases in model_metadata.json.

Canonical dataset: intestine. Full predictions and metrics: experiment-artifacts.

Numeric input values are the identical canonical float32 X shared by Linear, MLP, PCA and scGPT. No additional normalization, bins, HVG or sequence cropping.

Load using the family factory and metadata in the source repository. The safetensors files are native state dictionaries and do not imply AutoModel compatibility. Preserve gene order, vocabulary and class order. Historical initialization checkpoints were not captured; no newly sampled initialization has been substituted.

Collection: BioLLM mechanisms.

The private source repository is BogdanDidenko/biollm-mechanisms. Access to that repository requires permission; all files in this HF repository are public.

This release preserves the frozen historical experiment inputs and scientific endpoints. Future matched-initialization 1/2/3-layer panels are planned; their weights are not included. Previously inspected test results are exploratory. Historical 20L/12L models, backup aliases, temporary transfers and pairwise-distance caches are excluded.