fixed-14b / repair_notes.json
BRlkl's picture
Upload repaired Qwen3 causal LM checkpoint
3821b39 verified
{
"broken_model_id": "cemig-ceia/energygpt-14b-regulatorio",
"base_model_id": "Qwen/Qwen3-14B",
"tokenizer_source": "base",
"lm_head_source_shard": "model-00008-of-00008.safetensors",
"lm_head_dtype": "float32",
"warning": "This is a best-effort repair. If the original fine-tune updated lm_head during training, reusing the base lm_head may still underperform compared with the intended checkpoint."
}