ProtonPrat/anlp-a2-part2_lion

ANLP Assignment 2 — part2_lion

Final rerun checkpoint of the custom PyTorch decoder. Load with the course repository's src.part1.model.Transformer or scripts/infer.py; the export does not register a Transformers AutoModel architecture.

Measured results

Metric	Value
Training input positions	42,307,041
Total parameters	10,084,480
Final test perplexity	167.744570
Final test BLEU	0.656784
Generation examples	1000

Dataset: browndw/human-ai-parallel-corpus at b514ff64988d9e322fd81c5d70d69a38e78491f5.

One complete grouped-document corpus pass; test PPL is token-weighted, and continuation BLEU uses 1,000 fixed cached held-out prefixes.

Current W&B run · Report and code · Public raw generations

Saved files

• hf_export/model.safetensors, configuration and tokenizer: final inference export.
• final.pt and latest.pt: model, optimizer, RNG, cursor, config and data fingerprint for resume.
• best.pt: validation-selected weights, not a complete resume state.
• fraction_0.1.pt through fraction_1.0.pt: all ten optimizer milestone resume states.

from huggingface_hub import snapshot_download
directory = snapshot_download("ProtonPrat/anlp-a2-part2_lion", allow_patterns="hf_export/*")

Single-seed results describe this configuration. The report includes the three-point Lion sweep and paired BLEU uncertainty; no Sophia result is claimed. The implementation used LLM coding assistance and custom optimizer/decoding loops. Model confidence and reference overlap do not establish semantic quality.