You can deploy this model and start earning money today!

poolside-laguna-hackathon/
causal-discovery-research

Name: causal-discovery-research
Availability: InStock

poolside-laguna-hackathon

Views

No views yet

Poolside Laguna Causal Discovery

Hackathon submission scaffold for training and evaluating Laguna XS.2 on research-heavy causal discovery tasks.

This repo packages two connected research tracks:

xHSCIC: permutation-free conditional independence testing and method reconstruction from a paper-plus-code corpus.
cauchy: higher-order causal discovery with hypergraph structure and kernel interaction tests.

The training target is not a generic chat model. The environment asks Laguna XS.2 to recover research gaps, reconstruct methods, design decisive numerical experiments, and regenerate core Python implementations.

This initial public release is intentionally source-first. It includes the environment, reference corpora, eval configs, and release scaffolding now, and it reserves a clean artifacts/ surface for finalized eval runs, reports, and model outputs later.

Submission Snapshot

lab/environments/poolside_env/ contains the Prime/Verifiers environment used to standardize raw scientific material before prompting or judging.
lab/environments/poolside_env/reference/xhscic/ bundles the xHSCIC paper, implementation, and experiment assets.
lab/environments/poolside_env/reference/causal-higher-order/ bundles the cauchy paper scaffold, code, and smoke tests.
lab/configs/eval/ contains the Laguna XS.2 eval suites for both tracks.
scripts/publish_to_hf.py syncs this staged repo to Hugging Face.
scripts/stage_release_artifacts.py prepares future public eval artifacts in a cleaner release layout.

Why This Submission Is Interesting

It evaluates long-horizon research behavior instead of narrow benchmark QA.
It mixes paper understanding, code regeneration, and experiment planning in a single environment.
It covers two distinct causal-discovery regimes: permutation-free CI testing and higher-order interaction discovery.
It is set up to publish clean, machine-readable artifacts instead of dumping raw workspace outputs.

Reproduce The Current Snapshot

Install the environment and run the current eval suites:

bash

cd lab
prime env install poolside-env -p ./environments --plain
prime eval run configs/eval/laguna-xs2-causal-research.toml
prime eval run configs/eval/laguna-xs2-conditional-higher-order.toml

The published repo excludes local virtual environments, caches, build artifacts, and transient outputs/ directories.

Planned Public Artifacts

Finalized public releases will be organized under artifacts/:

curated eval snapshots with metadata.json and results.jsonl
model artifacts such as adapters, merged checkpoints, or quantized exports
short reports with plots, notes, and benchmark summaries
machine-readable manifests that index the public releases

This keeps the model card readable while making later updates predictable.

Release Layout

text

artifacts/
├── evals/       curated eval releases
├── manifests/   machine-readable indices for published artifacts
├── models/      adapters, merged checkpoints, quantized exports
└── reports/     plots, tables, short writeups, and benchmark notes

The helper below stages eval artifacts from local lab/outputs/evals/ into the public layout without copying over every transient log by default:

bash

python scripts/stage_release_artifacts.py --track laguna-xs2-causal-research --dry-run
python scripts/stage_release_artifacts.py --track laguna-xs2-conditional-higher-order --dry-run

When you are ready to publish staged artifacts, rerun without --dry-run and then sync the repo.

Updating The Hub Repo

Authenticate first if needed:

bash

cd lab
uv run python -c "from huggingface_hub import login; login(add_to_git_credential=True)"

Then create or update the Hub repo from this staged snapshot:

bash

cd lab
uv run python ../scripts/publish_to_hf.py \
  --repo-id poolside-laguna-hackathon/causal-discovery-research

publish_to_hf.py defaults to a model repo because that is the most visible artifact type in the hackathon org, but --repo-type dataset and --repo-type space are also supported.