Add merge script - run locally with HF_TOKEN set

Add merge script to create inference-ready model without ClippableLinear issues

Training in progress, step 954

Training in progress, step 954

Training in progress, step 500

Training in progress, step 500

Hardcode reduced sample limits (3500/4000) to fit A10G 8h timeout. Env vars keep old values on duplicated jobs.

Fix OOM: bs=1, accum=32, 1 epoch. Read sample limits from env (defaults: 3500/4000 for ~28K total to fit 8h)

Fix OOM: bs=2, grad_accum=8, 1 epoch. Hardcode reduced samples (5K/4K) for A10G 8h fit

Speed up training: batch_size=4, grad_accum=8, 1 epoch (fits in 8h timeout on A10G)

Fix: restore Gemma4ClippableLinear unwrapping + manual get_peft_model (PEFT throws ValueError on ClippableLinear modules)

Fix training script: remove unnecessary Gemma4ClippableLinear unwrapping, pass peft_config to SFTTrainer, add disable_tqdm=True for clean logs

Add debug test script to isolate SFTTrainer creation issue

Debug version with explicit logging and dataloader_num_workers=0

Fix OOM: reduce max_length to 1024, LoRA rank to 32, add PYTORCH_CUDA_ALLOC_CONF

Fix: Unwrap Gemma4ClippableLinear layers for PEFT LoRA compatibility

Fix: Use bf16 without 4-bit quantization to avoid Gemma4ClippableLinear PEFT incompatibility

Fix: Use AutoTokenizer instead of AutoProcessor for text-only training (avoids torchvision dependency)

Fix: Use tokenizer instead of full processor for Gemma4 text-only training

Add inference demo script

Add comprehensive README for Indian language finetuning

Add Indian language finetuning script

initial commit