training/TRAINING.md: add "Quick start — just run the .sh" subsection

training/TRAINING.md: add upfront "what the .sh launchers do" section

training/TRAINING.md: fix .sh / .py flag names so the recipe actually runs

Add training/TRAINING.md — end-to-end reproduction recipe

build_before_after_csv: --min-verbose-accuracy flag

trackio: post-hoc replay of train_metrics.jsonl into a HF Space dashboard

demo CSVs: add reward_advantage_vs_verbose + accuracy_delta_vs_verbose

demo: sample test input dropdown (per-task examples in CSV)

multistep: gradient checkpointing + tighter memory defaults

tasks_policy: long-context policy-compression tasks

hf_job_train: add ENABLE_THINKING env var (default true)

train_grpo: drop stale args.turn_limit reference at build_prompt_dataset call site

v3: multi-turn env, thinking tokens, cross-family Qwen->Llama, multi-step GRPO

profile_baseline: fix wrong TargetGeneration field accesses

profile: install hf_transfer (HF_HUB_ENABLE_HF_TRANSFER was set without the package)

profile_baseline: pass required max_output_tokens to generate_batch

tasks_tough: add 42 more tough scenarios + baseline profiler

tasks_tough: add 10 domain-classifier tough scenarios (seed batch)

eval: use merged v1+v2 task bank (same fix train_grpo.py already had)

Revert num_generations 10 -> 8 (must divide generation_batch_size=8)

Pre-launch fixes: disable Qwen3 thinking, strip think blocks, degenerate-short guard

GRPO: explicit temperature=0.9 top_p=1.0 (override Qwen3 defaults of 0.6/0.95 for rollout diversity)

Fall back from Qwen3.5 -> Qwen3 family (transformers==4.56.2 compat)

Fix v2 smoke failures: load merged task bank + newer transformers for Qwen3.5

v2 stack: Qwen3.5-2B agent/target, Qwen3.5-9B judge, hard tasks, additive reward

eval: --push-to-hub uploads eval JSONL to adapter repo under evals/

Revert agent loading to TRL + PEFT (Unsloth collides with frozen target)

Drop duplicate make_plots call; train_grpo.py now renders + uploads plots inline

Persist training artifacts: upload metrics + plots alongside adapter

Switch agent loading to Unsloth FastLanguageModel + fix padding side

Install openenv-core explicitly (our env imports it; --no-deps skipped it)

Adopt OpenEnv-official install pattern (unsloth_2048.ipynb)

Switch base image to python:3.12-slim + explicit torch 2.7 install

Default HF Jobs flavor to l40sx1 (48GB L40S)

Align HF Jobs deps with spaces_pipeline_env Colab stack

Bump HF Jobs base image to pytorch 2.5.1 (trl>=0.14 needs FSDPModule)

Fix bash -c invocation: add -- separator to stop hf CLI from eating -l -c as short flags

Fix hf jobs CLI: use --secrets plural + --detach, optional --push-to-hub

Initial commit: Prompt Golf environment for OpenEnv