fix(train): v5 — JSON preprocessor + dual-token for friend's account

fix(prompt): v4 — add explicit JSON examples for all action types

fix(train): v3 — fix fallback rewards, token budget, steps cap

fix(train): v3 — repair truncated JSON + earlier evals + fix GGUF path

fix(train): v2 — fix collapse root causes + best-ckpt tracking + GGUF export

fix(train): switch to l40sx1 — better availability, 300 steps fits $24 budget

fix(train): reduce steps to 250 — guaranteed to complete within $24.05

fix(train): reduce total_steps 400→350 to fit $24.05 budget

feat(train): switch to a100-large, reduce steps/tokens to fit $24 budget

feat(logging): full visibility into every training step

perf(train): optimise for H200 budget — 1000 steps in ~5h for $26

fix(train): also clone log_probs to escape inference_mode context

fix(train): clone prompt/completion tensors after inference_mode generation

fix(train): add for_inference/for_training toggles + set lora_dropout=0

fix(train): tokenize-free log-prob recompute + switch default model to Llama 3.1

fix(train): skip degenerate log-prob recomputes + cap default at 3500 steps

fix(train): allow gradients through compute_log_probs for current policy

fix(train): drop .to(device) calls on bnb 4-bit models — Unsloth places them already

fix: is_terminal bug in curriculum_basic + wire local judge into rollout_test

feat(training): local judge + parallel rollout + optimized batch config

chore(round2): full audit pass — harden rewards, patch frontend, tighten training

feat(round2): DB-grounded multi_domain task with 30-ticket curriculum stage

docs: reorganize .md files into docs/ and rewrite README for Round 2

feat(train): configurable model via TRAIN_MODEL env var

fix: Qwen3-8B training correctness — PPO formula, think tokens, model defaults

fix: Qwen3-8B-Instruct support — model name, think token stripping

feat: hackathon improvements — benchmark dashboard, GRPO training pipeline, grader fixes