Upload frontend/src/index.css with huggingface_hub

Upload frontend/src/App.jsx with huggingface_hub

feat: instant demo mode, inline results after attack, 3-option launch modal

feat: training results dashboard with stats cards, comparison table, and plots

feat: launch mode modal with demo vs live options and time estimate

feat: battlefield visualization — attacker/firewalls/agent with live reward graph

feat: fill in real results, fix openenv.yaml Space URL

feat: real continuous GRPO reward curve + 5-plot dashboard

feat: launch-mode modal (recorded vs live) + reframed success card

fix: shorten HF short_description to <= 60 chars

fix: HF Space URLs point to Jaswanth-K/Inject-Arena (the actual Space)

feat: wire frontend to real API, ship plots, multi-stage Docker build

Phase 7: replay traces for HF Space demo

Phase 7: replay traces for HF Space demo

feat: HF Space replay backend — trace store, /api endpoints, Docker, Cell 9

feat: 5 training plots (reward±std, KL/loss, completion stats) + Drive backup in Cell 8

fix: plots gitignore, trainer_state fallback for reward curve, Cell 8 --trainer-state arg

fix: Cell 8 — pull before plots, GH_TOKEN push, copy Drive logs

Phase 7/8: full README, video script, make_plots.py, Cell 8 implementation

feat: fill in Cell 7 for full training run (Phase 6)

fix: use huggingface_hub.login() before LoRA download to fix 401

fix: pre-download SecAlign LoRA before vLLM init to avoid 401 in subprocess

fix: dtype kwarg (torch_dtype deprecated), vLLM max_model_len=4096

fix: max_completion_length 512→128, firewall circuit-breaker

fix: drop BnB 4-bit, load attacker in plain bf16

fix: dtype kwarg + cast lm_head/embeds to bfloat16 to fix GRPO generate

fix: add torch_dtype=bfloat16 to prevent Float/BFloat16 mismatch in GRPO

fix: GRPO batch_size must be divisible by num_generations (1→4)

fix: skip Unsloth in GRPO trainer (grpo_accumulated_loss signature mismatch)

fix: stub GuidedDecodingParams for vLLM 0.19+ / TRL compatibility

Phase 5: training pipeline — client, GRPO trainer, eval, baselines (23 handcrafted attacks)

Phase 4: InjectArenaEnv + FastAPI server + Dockerfile + env tests (81 passing)

fix: load SecAlign before PG2 in Cell 3 to avoid vLLM CUDA init conflict

fix: add %cd /content/injectarena to cells 3+4 (lost after runtime restart)

fix: FirewallWrapper falls back to pg2 instance when llamafirewall scanner fails

fix: LlamaFirewall async handling + conditional AgentAlignment (TOGETHER_API_KEY)

fix: SecAlign fallback uses transformers+4bit instead of vLLM (T4 CUDA init conflict)

fix: PG2 label check — model returns LABEL_0/LABEL_1, not MALICIOUS/BENIGN

Phase 3: defense wrappers + Colab smoke/benchmark cells

Fix: widen requires-python to include 3.12

Notebook: hard-code REPO_URL to Inject-Arena GitHub

Phase 2: verifiers, embedding cache, reward function

Phase 1: schemas, safety filter, scenario bank

Phase 0: bootstrap