docs/08-timeline.md

08 — Hour-by-hour Timeline

48-hour sprint. Hard deadline: Apr 26, 5:00 PM IST.

Day 1 — Saturday, Apr 25

7:00–9:00 AM — Reporting & Opening (mandatory attendance)

• 7:00 AM check-in at Scaler Electronic City campus
• Breakfast (until 9 AM)
• Opening ceremony — listen for: Cursor credits, HF credits ($30?), final theme clarifications

9:00–10:00 AM — Setup

• Both team members:
  • Verify Colab account works, T4 reachable
  • Verify HF account, get on-site credits redeemed
  • Run hf auth login + hf jobs hardware to confirm t4-small access works (per opening-deck p.78-80)
  • Verify GitHub access
• Anurag: create empty clarify-rl HF Space (CPU basic, public, Docker)
• Kanan: create empty clarify-rl GitHub repo

10:00 AM–12:00 PM — Phase 1: Scaffold ✅ DONE

• All scaffold files in clarify-rl/:
  • openenv.yaml, pyproject.toml, Dockerfile, init.py, models.py, client.py, server/init.py
• ✅ Done. Move directly to Phase 2.

12:00–2:30 PM — Phase 2: Core environment (Anurag-led) ✅ DONE

• server/scenarios.py — procedural scenario generation
• server/user_simulator.py — rule-based Q→A
• server/clarify_environment.py — MCPEnvironment subclass with 3 tools + step_async
• server/app.py — FastAPI app via create_app(...)
• Local smoke test: uvicorn server.app:app + all 3 smoke scripts pass
• Lunch can be eaten while building (ladoos, biryani provided)

2:30–4:00 PM — Phase 3: Rubrics + per-step grader (Anurag-led) ✅ DONE

• server/rubrics.py — 5 Rubric subclasses, composed via Sequential + Gate + WeightedSum
• server/grader.py — per-step shaping reward + plan parser
• Wire rubric into ClarifyEnvironment.__init__
• Verified: oracle policy → ~0.89, random → low, blank plan → 0

4:00–5:30 PM — Phase 4: Deploy + baseline eval ← NEXT

• Push to HF Space, verify public access incognito
• inference.py — baseline eval script (Qwen2.5-1.5B-Instruct via HF Inference API)
• Run baseline on 100 held-out scenarios → save outputs/baseline.json
• Generate scenarios/eval_held_out.json (frozen seeds 10000-10099)

5:30–7:00 PM — Phase 5: Training notebook (Kanan-led, in parallel)

While Anurag builds env: Kanan forks openenv_wordle_grpo.ipynb (TRL official, opening-deck p.73) and adapts it.

• training/train_grpo.ipynb — fork of openenv_wordle_grpo.ipynb → swap env client to ClarifyClient, swap reward fn to ours
• Verify on Colab T4 first; switch to HF Jobs t4-small once smoke run passes
• Mock rollout with dummy env until real env is ready

7:00–9:00 PM — Phase 6: First real training run

• Dinner overlap (gulab jamun expected)
• Run 100 GRPO steps (Colab T4 first, HF Jobs t4-small once stable)
• Inspect reward curve: should trend up
• Eyeball 4 random rollouts (sample-inspection cadence per 06-training-plan.md) — reward up but degenerate generations = reward hacking, stop & fix
• If flat: debug (likely rubric or rollout parsing)

9:00–11:00 PM — Phase 7: Iterate

• If first run looked good: continue with 300 more steps
• If reward hacking observed: tighten rubric, restart
• Generate first preliminary plots

11:00 PM–1:00 AM — Phase 8: Overnight prep

• Start the LONG training run (600 steps)
• One person monitors checkpoints, the other sleeps in shifts
• Begin README.md draft (use 07-deployment.md as template)

Day 2 — Sunday, Apr 26

1:00–7:00 AM — Sleep shifts + monitoring

• Long run completes (~90min on T4) → if Colab kills session, resume from checkpoint
• Both should have at least 4-5 hours rest

7:00–9:00 AM — Phase 9: Final eval + plots

• Run trained model on held-out 100 scenarios
• Save outputs/trained.json
• Generate ALL 5 plots: reward_curve, loss_curve, per_task_bars, per_component, eval_dist
• Commit plots/*.png to repo

9:00–11:00 AM — Phase 10: Demo content (Kanan-led)

• Record demo screen captures (baseline vs trained traces)
• Edit 2-min video (OBS → ffmpeg → upload YouTube unlisted)
• OR write blog.md HF blog post (parallel option)
• Both ideally — videos pop in pitches

11:00 AM–1:00 PM — Phase 11: Polish + README

• Finalize root README.md with:
  • Pitch + headline metric
  • Embedded plots with captions
  • Result table
  • Architecture diagram (mermaid or ASCII)
  • All 4 deliverable links
• Polish blog.md if writing one
• Final commit

1:00–3:00 PM — Phase 12: Final validation sweep

• Run through docs/07-deployment.md validation checklist (every line)
• Test HF Space from incognito
• Re-run Colab notebook end-to-end on fresh runtime
• Verify all README links open correctly
• Fix any breakage

3:00–4:30 PM — Phase 13: Submit

• Fill Google Form with all 4 URLs
• Double-check each URL works from incognito
• Confirm submission

4:30–5:00 PM — BUFFER (do not skip)

• Re-verify submission record
• Take screenshots of submission confirmation
• Final coffee. Breathe.

5:00 PM — DEADLINE 🚨

No more changes accepted under any circumstances.

5:00 PM onward — Closing & networking

Team Split — Bhole Chature

Phase	Anurag	Kanan
1	scaffold review	scaffold review
2	core env (clarify_environment.py)	scenarios.py + user_simulator.py
3	rubrics.py + grader.py	unit tests for env + rubric
4	HF Space deploy	inference.py + baseline eval
5	bug-fix env on issues	train_grpo.ipynb (Colab)
6	monitor first training run	inspect outputs, sample logs
7	iterate on env/rubric	iterate on training config
8	overnight checkpoint mgmt	sleep shift A
9	sleep shift B	final eval + plots
10	README + repo polish	demo video + blog.md
11	architecture writeup	submission link audit
12	final validation sweep	final validation sweep
13	hits submit	screenshot + confirm

Both pair on critical risk items: training debug, validation sweep.

Anti-bus-factor Rule

Both team members must know how to:

• Push to HF Space + GitHub
• Restart the Colab runtime
• Run the env locally
• Run the eval script

If one is stuck/AFK at submission time, the other should be able to ship.

Energy Management

• Snack/coffee every 2 hours
• 5-min walks each hour
• No alcohol/Red Bull spirals
• Hydration > caffeine