Large Language Model Training Complete | Generated by AI
Training complete!
Status: DONE — step 130,000/130,000 Total time: 2,549.86 minutes (~42.5 hours) Val bpb: 0.780026 (minimum) Peak VRAM: 9,443.99 MiB Checkpoint: /home/lzw/.cache/nanochat/base_checkpoints/d12/model_130000.pt
Eval results (CORE metric: 0.1916):
hellaswag (10-shot): 0.43 winograd (0-shot): 0.62 winogrande (0-shot): 0.54 boolq (10-shot): 0.585 bigbench_cs_algorithms: 0.34 squad (10-shot): 0.27 coqa (0-shot): 0.26 agi_eval_lsat_ar (3-shot): 0.21 bigbench_dyck_languages: 0.17 bigbench_operators: 0.14 bigbench_repeat_copy_logic: 0.00 bigbench_language_id: 0.195
Sample generations look coherent but repetitive (typical for a 286M model). The checkpoint is saved and ready for chat or further eval.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋