grpo: add --eval_best_of (best-of-N eval for SFT and GRPO)

grpo: add --eval_best_of (best-of-N eval for SFT and GRPO)

GRPO v3_anchored: fix promote mode-collapse

GRPO v2_delta_phase reward + diversified states + tuned smoke schedule

GRPO v2_delta_phase reward + diversified states + tuned smoke schedule

Colab: HF Jobs orchestrator (full 400/200/3 run)

Update usage comment for full training

Bump default max_steps to 200

Fix notebook: unsloth-first imports, HF spaces clone, GPU check

Add Colab notebook for full GRPO training

Patch missing warnings_issued attr for transformers 5.x compat

Tighten pydantic pin to <2.11 to match mergekit ~=2.10.6

Pin pydantic <2.12 after editable install to fix mergekit compat

Add weave dep (another TRL 0.24 hard import in callbacks.py)

Patch TRANSFORMERS_CACHE in sanity check too

Patch TRANSFORMERS_CACHE for llm_blender compat

Add llm-blender dep required by trl 0.24 judges

Add mergekit dep required by trl 0.24 callbacks

Fix total_mem -> total_memory in GPU check

Add GRPO shell launcher with pinned deps

Add standalone GRPO training script

Fix GRPO smoke dependencies with llm-blender

Fix GRPO smoke dependencies with llm-blender

Fix GRPO training script import path

Add PYTHONPATH check to GRPO smoke launcher

Fix GRPO training script import path

Fix GRPO smoke dependencies with llm-blender

Pin TRL version for GRPO smoke job

Fix GRPO smoke job by installing mergekit

Fix GRPO smoke job torchao incompatibility

Add GRPO smoke job launcher

Add GRPO smoke job launcher

Add GRPO training script for DesignGym 2.0

Upload folder using huggingface_hub