Reza2kn/uncgpt-69-lfm2-langtag-fullpass-20260516

UncGPT 69 LFM2 Small Lang-Tagged Full Pass

LFM2-inspired UncGPT checkpoint trained on the 2026-05-15/16 quality-clean, language-tagged corpus.

Contents

• —inference_weights/lfm2small_langtag_step12399.safetensors: model-only checkpoint converted from the trainer checkpoint.
• —configs/stage_01_1k_lfm2small_live_qualityclean_langtag_3846_weighted_fullpass.yaml: training/model config.
• —tokenizer/uncgpt_live_qualityclean_langtag_3846_2026_05_15_multilingual_8192_byte.model: SentencePiece tokenizer with language tag user-defined symbols.
• —model/: minimal PyTorch model code used by the demo Space.
• —logs/: training logs.

Training Snapshot

• —Architecture: lfm2_small
• —Parameters: 69,791,680
• —Corpus: 3,846 conversations, 67,892 turns
• —Tokens: 12,696,040
• —Steps: 12,399 at 1024 tokens/step
• —Language tag format: conversation prefix like <lang:en> before <user>... / <uncle>... turns
• —Checkpoint SHA256 source trainer checkpoint: 231086b4266b943f7ee912f7b2e1fd3dfb614a8effba68b501a35b2bb7a61d61

This is an early small-data training run for pipeline/inference testing, not a finished product model.