Align train_reward_chart with v2 reward_log CSV; allow Python 3.10.

Fix Colab/Kaggle: install fastmcp after --no-deps openenv (import support_ops_env).

Split train and tool simulator modules; mastery curriculum and grader workflow nudge.

feat: default GRPO/eval to Qwen/Qwen3-1.7B + LoRA; disable Qwen3 thinking in chat template

fix: T4-safe 4bit compute, eval/audit resilience, async inference, plot_rewards CSV/table

Update notebooks and proof artifact tooling

Fix rollout logprobs and vLLM generation

Harden notebook vLLM compatibility checks

Trim notebook install dependencies

Fix OpenEnv and HF training flows

Docs: align with Hugging Face Hub + TRL official guides

Fix Colab TRL+vLLM: pin vllm<0.19, add tomli deps, guard train

Use SVG reward curve for HF Space Git (reject binary PNG)

Add DriftShield proof artifacts and generator script

Colab/Kaggle: auto GPU profile for bf16+vLLM on A100/L4

Constrain notebook runtime dependencies

Stabilize notebook dependency installs

Fix notebook package refresh and training runtime

Refresh Colab notebook for latest DriftShield trainer

Harden training parser and validation workflow

Improve DriftShield training prompts and logging

Diagnostic: track parse_ok_ratio per episode, dump health every 5 episodes

Fix flat-reward training: stronger SYSTEM_PROMPT with few-shot, longer completions, MAX_TURNS=16, schema sanity probe

Rename files to DriftShield: notebooks, server module, kernel folder

Rebrand: drop 'Support Ops' user-facing text; add DriftShield* aliases; tighten README

README: map DriftShield tasks to OWASP LLM Top 10 (2025)

Cleanup: drop legacy support-ops cruft now that DriftShield is the only collection

DriftShield Phase C: README pivot + driftshield_easy as default curriculum

DriftShield Phase B: 4 new grader components, 5 new penalties, negation-aware forbidden phrases

DriftShield Phase A: D1 collection, 4 runtime-failure tasks, ops.get_recommendation, recoverable schema drift

Explain GitHub import vs Kaggle Internet; friendlier gaierror in network check

Notebook UX: split pip vs imports; warn against Ctrl+C during DNS retries

Docs: clarify pip DNS failures; add pre-install network check cells

Add Kaggle training notebook, kernel metadata, and push script

fix: patch Qwen GRPO trainer setup for current TRL

Merge remote-tracking branch 'hf/main'

feat: GRPO training stack (Qwen3 + Unsloth opt-in, curriculum, eval, audit)

fix Gemma 4 path: pin transformers==5.5.4 and use AutoModelForImageTextToText + AutoProcessor in notebook

install transformers from git HEAD for Gemma 4 support (model_type=gemma4)

fix: rename SupportOpsToolEnv._reset -> public reset() for TRL environment_factory

add jmespath to [gemma] extra and Gemma 4 notebook install (required by TRL environment_factory tool parsing)

fix Section 5: load 4-bit quantized model explicitly, pass loaded model + processor.tokenizer to GRPOTrainer

fix notebook: pre-align torch family in install, add 4-bit quant to GRPOTrainer

fix: notebooks now warn instead of crashing when HF_TOKEN is not in Colab secrets

full sync: Qwen + Gemma 4 training layers, Colab notebooks, env core

pin torchvision/torchaudio in [gemma] so bitsandbytes forced torch downgrade does not leave them mismatched

relax torch upper bound so Colab's torch 2.10 is not downgraded

fix: map reward funcs to real grader keys + use .sync() adapter in notebooks

sync core env (Control Tower tasks + tool-call action schema)

clean up bytecode and egg-info noise