tjwjdgns011119/latentcomp-pisco-train_random_20_200

latentcomp-pisco-trainrandom20_200

COCOM/PISCO decoder-adapter model trained with randomly sampled input budgets from 25 to 200 chunks for retrieval-augmented question answering.

Included artifacts

• —Final model artifacts from pisco_train_decoder_random25_200_ddp_20260907_095358_pid976993/model
• —training_config.yaml with the training configuration

Trainer checkpoints, optimizer states, and random-number-generator states are intentionally not included.

Loading

This is a COCOM checkpoint with custom model code. Load it with trust_remote_code=True; access to the gated Llama 3.1 base model is also required.