Fix all stale 1.5B refs to 7B + LR/beta/completion fixes

Fix notebook: remove hardcoded 384, use defaults (640, lr=5e-5, beta=0.04)

Fix stagnation: LR 5e-5, beta 0.04, default 640

max_completion=640 temp=0.9 fix truncation

num_gen=4 for T4 + bump steps to 340

3-level

Make unsloth import lazy

3-level curriculum + 7B + reward fixes

Add training evidence: curriculum results, plots (LFS), parse_job_log helper

changed model

balanced

fixed mismatches

hf job

improved

python notebook

protection against reward hacking improved

training script

feat: implement GRPO training script with environment health checks and structured reward functions for bail assessment

removed pass test file

----

feat: implement GRPO training pipeline for bail assessment model and update README credits

feat: implement dataset loader, environment, and GRPO training pipeline for undertrial bail prediction

modified

implemented

Fix A3 (OOM eval), B9 (NDPS eligibility), B3 (direction-gated computation bonus), A8-pt2 (episode_id case lookup)

Fix 8 compliance gaps: repeat-action dedup+cache, min-steps hard block, criminal history tool (12th action), efficiency removed from training formula, circular import cleaned, yaml formula synced

Reward overhaul: add compute_reasoning_quality (anchoring+arithmetic+specificity+consistency), parity-grounds penalty, reduce outcome 40%->30%, add 10% reasoning quality signal

Fix 5 bugs: inference mode reset, step_counts in curriculum, adapter-only save (x3), DEMO001 false defence claim, episode_id in /reset

import fixed

Fix B1-B4: add 4 actions to openenv.yaml, export from __init__/client, fix reward range, remove global random.seed

Fix ACTION_MAP gaps: add 4 new tools to REST + WebSocket handlers; remove StepResult import collision

Add 4 missing actions: read_submissions, assess_flight_risk, check_case_factors, apply_proportionality (fixes 4.3d/e/g/h/i)

Fix model mismatch: openenv.yaml now correctly declares Qwen2.5-7B-Instruct

feat: implement core UndertriAI OpenEnv training environment with tool dispatch and reward logic

Fix all audit gaps: custody neutral, parity-first bias, skip penalty 0.40, statutory process reward, /observation endpoint, reset() timeout, drift determinism

Fix 3 teammate-caught crashes: statutory/bias wrong arg types in trainer, env.state() in WebSocket

Fix 5 audit gaps: conditional bail, action history, efficiency reward, train/val split, env API routing

Fix 6 vulnerabilities: /state crash, reward clamp, condition reward, XML exploit, tool-skip bypass, timeout enforcement

Fix demo: remove shuffle for deterministic seed=0; update stage 2+4 memos to match real seed=0 GT outcomes

Add seed param to /reset: demo pins to seed=0 per stage for consistent known episodes

Fix crash: call super().__init__() so self.rubric is set before _reset_rubric()

OpenEnv compliance: proper base class, SUPPORTS_CONCURRENT_SESSIONS, state @property, updated openenv.yaml

Rewrite README as full hackathon submission doc with env design, reward formula, curriculum, training arch

Apply 3 critical fixes: unify reward with server/reward.py, before/after eval + results.json, generation inspection callback

Fix install URL in training script to point to real HF Space

Add real dataset: 4 curriculum stage JSONL files (5MB total, 1200 HC judgments)

Fix demo: stage-aware memo submissions — correct recommendation per case type

Add 4 rich demo episodes (all stages incl. bias+BNSS); demo UI cycles random stages

Serve interactive demo UI at root route instead of raw JSON

Add root GET route with API overview for HF Spaces landing page