kothasuhas/dl_alchemy_a6_generalization

Assignment 6: generalization

The released Assignment 6 files for Deep Learning Alchemy, mirrored from the course Modal volume. These are raw files, like the course DCLM release; use hf download, not datasets.load_dataset.

export A6_REFERENCE_ROOT="$PWD/data/a6_generalization/v1"
uv run hf download kothasuhas/dl_alchemy_a6_generalization --repo-type dataset --local-dir "$A6_REFERENCE_ROOT"

Run from the assignments repository after its usual uv sync. No Hugging Face account or token is required. The complete bundle is about 2.25 GB. The A6 starter reads A6_REFERENCE_ROOT; call training.train(cfg) to train on your own CUDA GPU.

• splits/target_30k.npy: 30,000 training chunks; splits/validation.npy: 1,000 validation chunks. Both are uint8 arrays with 1,024 token IDs per chunk. Fine-tuning subsets are nested prefixes of the training array.
• vocab.json: the 19-token character vocabulary, including BOS and EOS. Cast IDs to torch.long for training.
• checkpoints/{related,dclm,noise}.pt: the three 10B-token initializations. Use Config(initialization='related'), 'dclm', or 'noise'. The starter loads DNA/noise weights in full; DCLM transfers the transformer body into fresh DNA embeddings/head. Full fine-tuning updates all parameters.
• The remaining checkpoints/, references/, ensemble_logits/, and histories/ files are the supplied baseline/reference artifacts. manifest.json records the exact files, sizes, SHA256 checksums, and W&B sources; SHA256SUMS supports sha256sum -c SHA256SUMS.

The target DNA is the exact course subset of prepared OpenGenome2. It uses case-normalized nucleotide/IUPAC characters, packed FASTA sequences with EOS and a BOS prefix, and an IID chunk split. Validation is not species-held-out. See that dataset card for preparation and attribution. The upstream OpenGenome2 DNA data is released under Apache 2.0 (Brixi et al., 2025). The DCLM initialization uses the course DCLM data linked above.