Hugging Face's logo

Spaces:
shreayas
/
ledgershield-controlbench
Running

App Files Community
ledgershield-controlbench / docs /openenv-hackathon-alignment.md
shreayas's picture
shreayas
Clarify Exquisite notebook as flagship submission path
3143f5d verified
preview code
|
Raw
History Blame Contribute Delete
11.7 kB

LedgerShield OpenEnv Hackathon Alignment

This document checks the current LedgerShield repository against the OpenEnv Hackathon India 2026 judging criteria and minimum submission requirements. It treats the project as two connected but separate training surfaces:

  • the original OpenEnv-connected SFT benchmark proof, and
  • the additive Exquisite environment-in-the-loop post-training layer.

The goal is simple: make it easy for a judge to verify that the repository contains a novel environment, a coherent reward and training pipeline, real before/after learning evidence, and a clear story.

Executive Verdict

LedgerShield aligns well with the strict submission guidance.

The repository already satisfies the non-negotiables:

The main repo improvements added for alignment are:

Recommended Training Notebook URL

If the submission form allows only one public Training Run Notebook URL, use the Exquisite notebook:

Why this is the best single-link choice:

  • it is the flagship environment-in-the-loop training story
  • it contains the clearest reward-improvement evidence for judges
  • it directly represents the stronger GRPO result that reaches 0.6606 mean score against a 0.6627 teacher
  • it still sits on top of the original SFT proof, which remains available as supporting baseline evidence

Supporting baseline link:

Judging Criteria Mapping

Criterion Weight LedgerShield evidence Verdict
Environment Innovation 40% POMDP enterprise AP fraud world, ASHTG formalism, calibration-gated authority, institutional memory, sleeper-vendor attacks, deterministic decision falsifier, certificate-required track, 9 official tracks Strong
Storytelling 30% README narrative, problem framing, pitch deck link, consolidated docs, original SFT report, Exquisite deep-dive report, dashboard, mini-blog source Strong after README tightening
Showing Improvement in Rewards 20% Original A10G SFT loss and reward plots, baseline-vs-trained comparisons, Exquisite GRPO reward curves, teacher-gap closure, policy ladders, safety frontier, per-case deltas Strong
Reward and Training Script/Pipeline Setup 10% Original TRL SFT script + launcher + Colab, additive self-play -> environment execution -> falsifier -> GRPO -> DPO scripts, coherent reward decomposition, artifact inventories Strong

Minimum Submission Requirements

Requirement Evidence in repo Status
Use OpenEnv latest release and framework ../openenv.yaml, ../pyproject.toml, ../requirements.txt, FastAPI app wiring, reset/step/state environment contract documented in DOCUMENTATION.md Satisfied
Working training script using Unsloth or Hugging Face TRL Original path: ../training/ledgershield_trl_training.py, ../training/launch_hf_a10g_qwen_job.py Satisfied
Ideally a Colab notebook judges can rerun Original path: ../training/LedgerShield_OpenEnv_TRL_Training_Colab.ipynb; additive path: ../training/exquisite/LedgerShield_Exquisite_Training_Colab.ipynb Satisfied
Evidence that training actually happened ./training-report.md, ../artifacts/trl-openenv-hf-a10g-qwen-rich/, ../artifacts/exquisite-training/ Satisfied
Loss and reward plots from a real run Original plot pack under ../artifacts/trl-openenv-hf-a10g-qwen-rich/plots/, Exquisite plot pack under ../artifacts/exquisite-training/plots/ Satisfied
Short writeup, blog, video, or slide deck linked from README Public pitch deck link in ../README.md, plus linked docs and mini-blog source Satisfied
Environment pushed to a Hugging Face Space Linked in ../README.md as Hugging Face Space Satisfied
README motivates problem, explains env, and shows results ../README.md Satisfied
README links to the Space and extra materials ../README.md Satisfied

Original SFT Benchmark Path

This path is the baseline proof that the team really trained against the environment. It remains important supporting evidence even if the single submission-form notebook points to the Exquisite run.

What it proves

  • live environment trajectory collection
  • TRL SFT on executable LedgerShield plans
  • held-out improvement over random, naive, and base-model baselines
  • committed loss/reward/safety/certificate plots
  • a judge-rerunnable Colab notebook

Primary files

Key numbers

  • Base Qwen 0.5B: 0.1283
  • SFT Qwen 0.5B: 0.4394
  • Held-out parse success: 1.0000
  • Held-out unsafe release: 0.0000

This path alone already satisfies the minimum training requirement well.

Additive Exquisite Training Path

This path is the recommended single-link notebook submission because it gives judges the strongest end-to-end training story in one place: self-play, environment execution, deterministic reward, GRPO improvement, and final plots.

What it proves

  • the environment is usable as a post-training surface, not just an evaluation benchmark
  • self-play candidate generation produces a nontrivial quality distribution
  • deterministic reward and falsifier scoring can rank those candidates
  • GRPO improves the same model family from 0.4394 to 0.6606
  • the additive pipeline preserves 0.0000 unsafe release and 1.0000 parse success

Primary files

Key numbers

  • SFT Qwen 0.5B: 0.4394
  • GRPO Qwen 0.5B: 0.6606
  • Teacher: 0.6627
  • GRPO teacher-gap closure: 99.6%
  • GRPO unsafe release: 0.0000
  • GRPO parse success: 1.0000

Honest caveats

  • The completed SFT Qwen 1.5B artifact is a fast-profile scaling run on a smaller held-out slice, so it should be described as a scaling signal rather than as a flagship apples-to-apples comparison.
  • The repo should present GRPO as the flagship additive result. DPO is implemented and complete, but it is not the best final policy.

These caveats do not weaken the core submission. They simply make the storytelling more honest and credible.

Why The Reward Story Is Coherent

The reward and evaluation setup is one of the strongest parts of the repository:

  • the environment uses shaped reward plus terminal rubric reward rather than a single brittle binary success bit
  • the rubric includes certificate quality, control satisfaction, institutional utility, and safety-sensitive penalties
  • the additive training layer uses deterministic environment outcomes and falsifier signals, not an unrelated offline heuristic
  • the best improved policy does not gain score by taking unsafe shortcuts

The most judge-relevant evidence is visible in:

Recommended Judge Reading Order

For a fast 3-to-5 minute evaluation pass:

  1. ../README.md
  2. ./training-report.md
  3. ./exquisite-training-layer.md
  4. ./exquisite-visual-analysis.md
  5. ../artifacts/exquisite-training/dashboard/index.html

For a deeper technical pass:

  1. ./DOCUMENTATION.md
  2. ../training/README.md
  3. ../training/exquisite/README.md
  4. ../openenv.yaml

Bottom Line

LedgerShield now presents a strong two-layer training story that aligns with the OpenEnv Hackathon rubric:

  • a clear, runnable, OpenEnv-native benchmark
  • a real original TRL SFT training proof with rerunnable notebook and plots
  • an additive environment-in-the-loop GRPO layer that visibly improves behavior and rewards
  • a README and doc stack that points judges directly to the evidence