Upload training/train_grpo.py with huggingface_hub

v6: GRPO with --resume-adapter to start from SFT checkpoint

v5d: save best-epoch checkpoint during SFT

v5c: add --min-investigation-steps filter for submit-only with longer contexts

v5b: add --submit-only SFT mode

v5: remove 4-bit quant, variable-length traces, submit oversampling, GRPO exploration fixes

Upload folder using huggingface_hub

Fix: decode with skip_special_tokens=False to preserve <think> tags for stripping

Fix Qwen3 thinking mode + increase max_new_tokens: training/launch_hf_job.py

Fix Qwen3 thinking mode + increase max_new_tokens: training/eval_adapter.py

Fix Qwen3 thinking mode + increase max_new_tokens: training/train_sft.py

Fix Qwen3 thinking mode + increase max_new_tokens: training/train_grpo.py

Fix Qwen3 thinking mode + increase max_new_tokens: training/rollout.py

Fix Qwen3 thinking mode + increase max_new_tokens: inference.py

Add InvoiceGuard SFT and merge tooling

Add InvoiceGuard adapter eval job script

Sync InvoiceGuard code for GRPO training job

Sync InvoiceGuard code for GRPO training job

Sync InvoiceGuard code for GRPO training job

Sync InvoiceGuard code for GRPO training job

Sync InvoiceGuard code for GRPO training job

Sync InvoiceGuard code for GRPO training job

Sync InvoiceGuard code for GRPO training job

Sync InvoiceGuard code for GRPO training job

Sync InvoiceGuard code for GRPO training job

initial commit