ho22joshua/hep-signature-backgrounds

HEP Signature Backgrounds

This dataset contains generated high-energy-physics examples for mapping signal-region final-state signatures to Standard Model background compositions.

Contents

• —hep_sft/train.parquet, hep_sft/val.parquet, hep_sft/test.parquet: supervised fine-tuning splits.

Answer Schema

This is a conversational prompt-completion SFT dataset, following the Hugging Face/TRL convention:

{
    "id": "sigbg_e0_mu0_tau_h0_gamma0_b1_j0_met1_dominant_irreducible",
    "task": "dominant_irreducible",
    "final_state": {
        "electrons": 0,
        "muons": 0,
        "hadronic_taus": 0,
        "photons": 0,
        "b_tagged_jets": 1,
        "light_jets": 0,
        "has_missing_transverse_momentum": true,
    },
    "prompt": [
        {"role": "system", "content": "..."},
        {"role": "user", "content": "..."},
    ],
    "completion": [
        {"role": "assistant", "content": "<answer>...</answer>"},
    ],
}

prompt is supplied to the model; completion is the supervised target and is withheld during generation. id, task, and final_state are evaluation/filtering metadata; training code should select only prompt and completion. No answer-bearing metadata or duplicate messages field is included.

The user asks one of two separate questions about a reconstructed final state:

• —the dominant irreducible Standard Model background; or
• —the dominant reducible Standard Model background.

The assistant response contains exactly one human-readable process label, enclosed in <answer>...</answer>. It does not request or provide structured dominant/irreducible/reducible lists.

Intended Use

The dataset is intended for research on HEP-focused instruction tuning, validation, and reward modeling. It is not a replacement for detector simulation, collaboration-approved background estimates, or expert analysis review.

Generation

The default dataset path in the associated training repository generates synthetic signature_to_backgrounds examples from reconstructed particle combinations and validates outputs against a stable process catalog.

Training code: ho22joshua/hep-posttraining

License and Provenance

Please verify all upstream source and generation constraints before redistributing derived datasets or using them commercially.