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

Dockerfile: enable openenv web UI at /web (fixes Space 404)

docs: add multi-step training curves to README + BLOG_POST

BLOG_POST: clarify ambiguous "80% of 94-token human-prompt accuracy" in hook

Remove stale root TRAINING.md

BLOG_POST: stronger 3-paragraph hook, move Prior Work above the fold, escape unsafe single tildes

README: use escaped \~ for single-tilde approximations

docs: refresh BLOG_POST stale 87-task numbers; finish README ~ cleanup

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

README: replace ~ with ≈ in intro to fix accidental strikethrough

README: stronger intro

docs: add Prior Work section + replace Training/thinking-mode A/B with multi-step setup

docs: surface multi-turn in How-it-works, defer training-process notes

docs: consolidate Results section — single master table + per-category + examples

docs: add multi-step variant section to README + BLOG_POST

Update BLOG_POST.md

demo(new-tab): expose the raw chat-templated string sent to target

demo(new-tab): also run target with verbose description

demo: add 'Try a new task' tab

BLOG_POST: drop 37× MSN policy compression callout

BLOG_POST: drop policy-tasks-as-headline-workload framing

docs: add demo screenshots for blog post

BLOG_POST: integrate user revisions

BLOG_POST: drop 37× compression claims

BLOG_POST: full rewrite — research framing, 10 sections, citations, image placeholders

build_before_after_csv: --min-verbose-accuracy flag

demo: filter tasks dead on target (verbose=0 AND trained=0)

README: add Scorers section (21 scorers grouped by family)

docs: drop the misleading 37× compression anecdote (0-accuracy task)

README + BLOG: drop multi-step + Llama-self mentions (in-progress runs)

Add BLOG_POST.md (HF blog draft, hackathon framing)

remove untested Colab notebook + link training/ folder in README

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

demo CSVs: add reward_advantage_vs_verbose + accuracy_delta_vs_verbose

README: rewrite for hackathon submission — links-first, plots inline, kill verbose sections

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

demo: apply chat template to target (fix rambling completion-mode outputs)

multistep: gradient checkpointing + tighter memory defaults

space-demo: bump to gradio 5.x; drop jinja2/hub pins

space-demo: pin Python 3.11 + exact gradio + jinja2<3.1.5 (defuse 3.13 fallout)

space-demo: pin huggingface_hub<1.0 + cap transformers/gradio majors

space-demo: add audioop-lts for Python 3.13 (pydub fix)

space-demo: fix short_description length (HF Spaces 60-char cap)

space-demo: bundle for HF Spaces Gradio demo

inference.py: enumerate all task banks, drop stale TASK_NAMES gate

ui: local Gradio demo app — verbose / untrained / trained side-by-side

tasks_policy: long-context policy-compression tasks

notebooks: minimal Colab training demo