LiitLLM v1.0 — model weights

32.8M-parameter Taglish language model, trained from scratch (60,000 steps).

• liitllm-taglish.pt — the model trained on Taglish-filtered text (seed 1)
• liitllm-control-unfiltered.pt — the identical control trained on unfiltered text (seed 1)
• tokenizer.json — needed by both
• SHA256SUMS.txt — checksums

Weights only (no optimizer state), loadable with torch.load(..., weights_only=True).
Run: python -m liitllm.sample --ckpt liitllm-taglish.pt --tokenizer tokenizer.json --prompt "Kahapon nag-"
All four runs, both seeds per arm: https://www.kaggle.com/datasets/johnandreimartinez/liitllm-taglish