Set HF_HOME and HF_DATASETS_CACHE for writable caches on Space

Set TORCHINDUCTOR cache dirs before torch import (HF Space no passwd uids)

Work around torch mega-cache double registration on Space imports

Enhance TorchGenerator to support optional subfolder for model loading

Post-training push to Hugging Face Hub: --hub-model-repo and pilot query params

Document HF token via Space secrets, ENV.example, and operational curl for pilot and push_to_hub

Add 8-bit Adam + gradient checkpointing for bf16 training to fit 1.7B on L4

Add --bf16 flag and expose it via /train/pilot for bigger models

Force Qwen3 chat template to disable thinking mode via tokenizer patch

Fix reward_func reading wrong env layer for tool_calls counters

Nuclear reward shaping: silent rollouts get flat -3, active rollouts floor at +1

Crush-signal reward shaping to force exploration over silence

Fix reward collapse: floor exploration loss, bonus per-call and per-terminal

Live training visualization + aggressive reward shaping to prevent 'do nothing' collapse

Add compact training summary endpoint

Use Qwen3 model for TRL OpenEnv tool parsing

Adapt GRPO config to installed TRL API

Deploy ReleaseOps Arena GPU Space