harshbhatt7585/tinygroot-sft-mtp-heads-20260527-2

tinyGroot checkpoint uploaded from the training pipeline.

Checkpoint

• Step: 974
• Tokenizer: nanochat
• Layers: 12
• Hidden size: 768
• Heads: 6
• MTP heads: 2
• Max sequence length: 2048
• Source run: sft-mtp-heads-20260527-2

Files

• model.pt: model state_dict (mmap+weights_only=True friendly).
• optimizer.pt: optimizer and grad-scaler state (only needed to resume training).
• meta.json: step, model config, tokenizer type, and original training args.
• tokenizer_hf/tokenizer.json: tokenizer used by this checkpoint.

Load this checkpoint with the tinyGroot codebase, not the Transformers AutoModel API.