sparel/herg-curation-v2

hERG curation v2 QLoRA adapter

A PEFT LoRA adapter for converting supplied public human KCNH2/hERG assay and activity context into a flat JSON evidence card. It is an adapter-only release: load it with the compatible Qwen2.5-7B-Instruct base model.

Intended use

Use the adapter to create reviewable structured evidence cards from supplied records. It does not predict hERG potency, establish cardiac risk, infer clinical outcomes, or replace scientific review.

Safety boundary

The qsar_usable and exclude_reason fields are extraction outputs, not authoritative scientific or dataset-admission decisions. Recompute numerical-QSAR eligibility independently from the source fields before admitting a record to a QSAR dataset.

This adapter is not for clinical advice, patient-specific assessment, autonomous compound triage, safety conclusions, development decisions, or regulatory use.

Base model and adapter

• —Base model: Qwen/Qwen2.5-7B-Instruct
• —Base-model revision: a09a35458c702b33eeacc393d103063234e8bc28
• —Format: PEFT LoRA adapter, not merged base-model weights
• —PEFT version: 0.20.0
• —LoRA rank: 16; alpha: 32; dropout: 0.05
• —Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Tokenizer compatibility

Use the tokenizer and chat template supplied by Qwen/Qwen2.5-7B-Instruct at the pinned revision. This adapter does not add tokenizer tokens. Render the supplied record as plain text using the field order below.

Training data

Training used ChEMBL 37 public hERG/KCNH2 assay and activity records for target CHEMBL240. The web-services snapshot was retrieved on 2026-07-30: activities were written at 11:24:05 CEST, and assays plus the retrieval manifest completed at 11:24:30 CEST.

Query target: CHEMBL240. Source-file SHA-256 values: activities e7d76304534b3e2110c9b026cdb54ce2c32391b8b3ef8e2265dfcb08667c480a; assays 8f927acca9f2d61b07d91cfd1216853ba88ebe245b6e6ae40285987ae4ac4957; target f375681f4e0d8f2505cc376c0f552530f20eec643331f26f1eaec0f695f75c04. The adapter release contains no raw records, processed corpus, review material, generated outputs, or base-model weights.

ChEMBL data are available under CC BY-SA 3.0. This card is not legal advice; users are responsible for evaluating applicable attribution and downstream licensing obligations.

Deterministic admission contract

Final numerical-QSAR admission must be computed outside the model:

admit = endpoint == "IC50"
    and relation == "="
    and value is numeric and value > 0
    and units in {"nM", "uM"}
    and data_validity_comment is empty
    and canonical_smiles is non-empty

Input contract

Render one name: value line per field in this exact order:

assay_description
assay_type
bao_label
standard_type
standard_relation
standard_value
standard_units
standard_text_value
activity_comment
data_validity_comment
potential_duplicate
canonical_smiles

Use this instruction before the fields:

You are curating human KCNH2/hERG (CHEMBL240) records.
Return JSON only: one flat object, no markdown, no explanation, no nesting, and no keys beyond this exact schema:
{OUTPUT_SCHEMA}

Rules: copy supported source values exactly; use null for unknown scalar fields; `qsar_usable` is true only for an exact, positive IC50 in nM/uM with no data-validity flag and a non-empty canonical SMILES; otherwise false and populate `exclude_reason` with the applicable policy reason(s).

Assay/activity record:
{fields}

Output schema

{
  "assay_type": "string",
  "target_name": "string",
  "target_id": "string",
  "target_confidence": "high | unknown",
  "organism": "string | null",
  "cell_line": "string | null",
  "endpoint": "string",
  "relation": "= | < | > | <= | >= | unknown",
  "value": "string | null",
  "units": "string | other",
  "qsar_usable": true,
  "exclude_reason": ["string"],
  "curation_notes": "string"
}

Limitations

The adapter can produce valid JSON while making semantically unsuitable extraction decisions. Validate output syntax and source provenance, recompute hard eligibility rules deterministically, and route ambiguous records to qualified review.

Files and integrity

• —adapter_model.safetensors SHA-256: 4caf4813197c7c18c86079cadfabdbc88644395c7fc6ab8c1cb019befb8da04a.
• —adapter_config.json names the base model and pinned revision.
• —LICENSE applies Apache-2.0 to this adapter package.

Maintenance

Maintainer: Serge Parel.

Changelog

1.0.0 - 2026-09-06

Initial adapter-only release candidate.