| { |
| "comment": "HYP-36: Champion enhancement test. Hyp-32 (soft_800k_20ep, bs=512) is the current best at AUROC@50=0.7901 (+0.005 vs Hyp-14, +0.011 vs TCRdist) but the marginal lift over Hyp-14 says the 'more epochs' axis is essentially saturated. The remaining cheap knob inside the InBatchTcrdistSoftCollator pipeline is BATCH SIZE: doubling bs from 512 to 1024 doubles each anchor's pool of in-batch candidates (and the soft-positive set scales with that pool). This sharpens the contrastive signal per gradient step in a way that more epochs cannot. Single-variable change from Hyp-32: batch_size 512 -> 1024. Step count halves (827k/1024 ~= 808 steps/epoch * 20 = ~16,200 steps, comparable to Hyp-14's compute budget) but each step sees 4x more pairwise relations. Distinct save name 'SCEPTR6L_AC_soft_800k_20ep_bs1024' so we can compare directly against Hyp-32 in the same benchmark sweep.", |
| "training_delegate": { |
| "class": "ClTrainingDelegate", |
| "initargs": {} |
| }, |
| "model": { |
| "name": "SCEPTR6L_AC_soft_800k_20ep_bs1024", |
| "path_to_pretrained_state_dict": null, |
| "token_embedder": { |
| "class": "CdrSimpleEmbedder", |
| "initargs": {} |
| }, |
| "self_attention_stack": { |
| "class": "ScaledSceptrEncoderWithProjection", |
| "initargs": { |
| "embedding_dim": 29, |
| "num_layers": 6, |
| "embed_dim": 128, |
| "num_heads": 8, |
| "dropout": 0.1 |
| } |
| }, |
| "mlm_token_prediction_projector": { |
| "class": "AminoAcidTokenProjector", |
| "initargs": { |
| "d_model": 128 |
| } |
| }, |
| "vector_representation_delegate": { |
| "class": "ClsVectorRepresentationDelegate", |
| "initargs": {} |
| }, |
| "trainable_model": { |
| "class": "ClTrainableModel", |
| "initargs": {} |
| } |
| }, |
| "data": { |
| "training_data": { |
| "dataset": { |
| "class": "TcrDataset", |
| "initargs": {} |
| }, |
| "dataloader": { |
| "class": "SingleDatasetDataLoader", |
| "initargs": { |
| "batch_size": 1024, |
| "num_workers": 4 |
| } |
| }, |
| "csv_paths": [ |
| "tcr_data/preprocessed/tanno/train.csv" |
| ] |
| }, |
| "validation_data": { |
| "dataset": { |
| "class": "TcrDataset", |
| "initargs": {} |
| }, |
| "dataloader": { |
| "class": "SingleDatasetDataLoader", |
| "initargs": { |
| "batch_size": 1024, |
| "num_workers": 4 |
| } |
| }, |
| "csv_paths": [ |
| "tcr_data/preprocessed/tanno/test.csv" |
| ] |
| }, |
| "tokeniser": { |
| "class": "CdrTokeniser", |
| "initargs": {} |
| }, |
| "batch_collator": { |
| "class": "InBatchTcrdistSoftCollator", |
| "initargs": { |
| "frac_dropped_tokens": 0.2, |
| "prob_drop_chain": 0.5, |
| "training_csv_path": "tcr_data/preprocessed/tanno/train.csv", |
| "tau_tcrdist": 50.0, |
| "tcrdist_threshold": 200.0 |
| } |
| } |
| }, |
| "loss": { |
| "cross_entropy_loss": { |
| "class": "AdjustedCrossEntropyLoss", |
| "initargs": { |
| "label_smoothing": 0.1 |
| } |
| }, |
| "contrastive_loss": { |
| "class": "DotProductLoss", |
| "initargs": { |
| "temp": 0.05 |
| } |
| } |
| }, |
| "optimiser": { |
| "initargs": { |
| "n_warmup_steps": 2000, |
| "d_model": 128 |
| } |
| }, |
| "num_epochs": 20 |
| } |
