Drop GRPO temp to 0.3, bump max_new_tokens to 2048, add inference smoke test

Add inference-time retry loop with feedback for self-correction

Add in-context demonstration learning support

Fix make_plots.py to accept --wandb-run-id separately from --training-log-json

pytorch fixes

no unsloth

Retry torch.cuda.is_available() in fresh python procs; NVML ready != cuInit ready

Rewrite run_hf_job_7b.sh end-to-end: nvidia-smi GPU poll, clean phase structure

Drop venv approach; force-reinstall torch stack over base image to actually upgrade past 2.5.1

Phase 0: build fresh venv + restore CUDA poll + restore upfront huggingface_hub install

Install torchaudio alongside torch/torchvision to clear pinned-base-image conflict

Drop version pins on training extras and torch/torchvision install

Hard reset HF Job env: install torch+torchvision together at fixed versions

Upgrade torchvision/torchaudio alongside torch to fix nms op registration

Upgrade torch to 2.5+ in Phase 0; pin trl>=0.14 for GRPOTrainer

Make 7B HF Job Phase 0 resilient: poll for CUDA readiness, install hf_hub up front for trap

Use --env instead of --secrets KEY=VALUE for WANDB_API_KEY

Fix wandb login in 7B HF Job: verify env var present, use CLI not Python API

Add 7B variant of HF Job for stronger cold-start.

Add reduced-scope HF Job variant for time-constrained scenarios

Switch HF Job base image to pytorch:2.5.1-cuda12.4 to fix CUDA13 ABI bleed

Fix torch/torchvision ABI mismatch in HF Job Phase 0

Fix GitHub username in HF Job scripts: jester1177 -> melohub-xbit

Fix submit_hf_job.sh syntax: image as positional arg

Pre-submission HF Job patches: Python version compat + skip torch reinstall

Add HF Job training pipeline: persistence-aware run script, judge-facing demo notebook, baseline JSON output

Re-apply training prep work on top of validated Layers 1-5 base

Initial commit: MutantHunter — RL env for mutation-score-rewarded test generation