Mar 24

1. 13:33

   GGUF export only script

   lokegud
   Verified signature

   b6c77c6
2. 09:02

   Stage 2 only: tool prediction on personality LoRA, 3 epochs, batch=2

   lokegud
   Verified signature

   f1bc339
3. 02:00

   Fix: disable eval (OOM at step 40 — model uses 77/80GB). Train loss converging fine.

   lokegud
   Verified signature

   467c863
4. 01:33

   Fix: DataCollatorForSeq2Seq for proper label padding, removes eval crash at step 40

   lokegud
   Verified signature

   4e9b79b
5. 00:57

   Fix: extract text tokenizer from VLM Processor to avoid image_utils error

   lokegud
   Verified signature

   9dd474e
6. 00:48

   Fix: pre-tokenize with input_ids+labels instead of relying on SFTTrainer auto-tokenization

   lokegud
   Verified signature

   c3edb01
7. 00:41

   Fix: use messages format directly instead of text column for SFTTrainer 5.x compat

   lokegud
   Verified signature

   b34273c
8. 00:33

   Fix: remove_unused_columns=False for transformers 5.x, warmup_steps instead of warmup_ratio

   lokegud
   Verified signature

   5846d17
9. 00:03

   Upload train_hf_jobs.py with huggingface_hub

   lokegud
   Verified signature

   7528128
10. 00:03

    initial commit

    lokegud
    Verified signature

    9605eda