Keep mx_obs after episode done, prevent reset_multi first error

Fix API_BASE_URL to correct HF inference /v1 endpoint

Fix HF Inference API URL to include model path /v1

Fix API_BASE_URL - remove duplicate /v1

Switch to HF Inference API + zephyr-7b (no provider needed)

Switch to Phi-3.5-mini-instruct (free HF router)

Switch to Llama-3.2-3B-Instruct (supported by HF router)

Add timeout+stderr logging to LLM calls, reduce max_tokens to 64

Log to stderr + gunicorn capture-output for HF diagnosis

Cache bust + debug env print for HF token diagnosis

Fix Dockerfile MODEL_NAME to Qwen2.5-0.5B

Fix checkpoint path + default model to Qwen2.5-0.5B for HF API

Default to multi-agent mode on load

Track PNGs with LFS

Run 3: Qwen2.5-0.5B GRPO 500 steps | reward 0.96→1.75 | conflict 0.375→0.1875 | dashboard chart | agents wired

add /game route for game mode

add game mode files from remote

final submission: real reward curve from training logs

fix: clean up auto_multi, set mx_obs=None on done

remove filter: let LLM produce natural conflicts for LLM vs LoRA comparison

fix: use correct steps/seed per task on mode switch

debug max_steps + safety counter in runAllMulti

fix: await reset before runAllMulti loop

fix: auto-reset multi env when switching to multi-agent mode

debug: log filtered proposals before step_multiagent

fix: enforce sensor ownership server-side, drop cross-agent proposals

fix: restrict LLM proposals to agent-type sensors, prevent S1 conflicts

auto-load easy task on page load

add lora checkpoints and training updates

exclude checkpoints from hf space

track png with lfs

fix: rewards, conflict detection, training comparison panel

Merge branch 'main' of https://github.com/smritis21/arya

Merge pull request #12 from smritis21/chathu01

Merge branch 'main' into chathu01

feat: improve UI visualization and training evaluation with smooth transitions and before/after metrics

Remove AUDIT_UPDATED.md and reward_curve.png

Fix: task difficulty counts, dashboard legend modal, training metrics, reward curve, /metrics/history endpoint

Merge pull request #11 from smritis21/chathu01

fix: resolve training stagnation by randomizing curriculum seeds and enabling real-time metrics logging

fix: evaluate() uses trained agents, grader uses agent instances in eval loop

feat: add format_state() + llm_propose() — structured ISR prompts with temperature=0.3 generation and JSON validation

feat: cap GRPO at 30 steps via GRPOConfig(max_steps=30) — correct TrainingArguments approach

fix: remove max_steps from train() call — TRL GRPOTrainer does not support it as a runtime arg

feat: fast demo mode — 20 episodes, 10 steps, GRPO capped at 30 steps, model saved to checkpoints/

fix: strip unsupported GRPOConfig args — Unsloth patched config only accepts minimal args

refactor: modularize GRPOConfig definition and implement robust keyword argument handling for GRPOTrainer initialization

refactor: improve model loading robustness and simplify simulation environment logic in train_colab.py

train collab changes

Person 3: unify AryaXEnv mode param + fix evaluate() to use trained agents