PLM-HalluBench: A Multi-level Benchmark for Evaluating Hallucinations in Protein Language Models

NeurIPS 2026 Evaluations and Datasets Track submission (double-blind).

PLM-HalluBench is a benchmark for evaluating hallucination in protein language models (PLMs) — outputs that look like proteins but violate basic biophysics. It is organised around a three-level taxonomy that separates sequence-, structure-, and function-level failure modes, and it pairs a Factual track (BPHS against a natural reference band) with a Faithful track (Foldseek top-20 annotation matching).

• 1,000 functional conditions (Pfam families × organism / function prompts)
• 5,000 natural reference sequences from UniProt / Swiss-Prot (5 homologs per condition)
• 75,000 scored generations from six PLMs across three tasks
• Full BPHS scoring: ESMFold pLDDT/pTM, 8-component S_seq, signed BPHS, factual flag, Foldseek recall, faithful flag

Files

File	Records	Description
conditions.jsonl	1,000	One JSON record per line with cond_id, four model-aligned prompt formats, and a 5-homolog reference_set. A legacy dict-of-dicts copy is also kept as conditions.json.
natural_references.jsonl	5,000	Natural reference sequences with full BPHS scoring (source = "swissprot")
generations.jsonl	75,000	Model-generated sequences with full BPHS scoring and (for Text-to-Seq) Foldseek results (source = "generated")
plm-hallubench-croissant.json	—	Croissant metadata file (Core + Responsible AI fields)

Full JSON schema is documented in Appendix E of the paper.

Models and tasks covered

Model	Architecture	Unconditional	Seq-to-Seq	Text-to-Sequence
ESM3	Multimodal ESM	✓	✓	✓
ProLLaMA	Instruction-tuned LLaMA	✓	✓	✓
ProtGPT2	Autoregressive (causal LM)	✓	✓	✓
EvoDiff	Diffusion	✓	✓ (motif scaffolding)	—
ESM-2	Masked LM	✓	✓	—
ProGen2	Autoregressive (causal LM)	✓	✓	—

Quickstart

from datasets import load_dataset

# Load conditions
conds = load_dataset("plm-hallubench/plm-hallubench", "conditions")

# Load natural references
refs = load_dataset("plm-hallubench/plm-hallubench", "natural_references")

# Load generations
gens = load_dataset("plm-hallubench/plm-hallubench", "generations")

# Headline metric: factual hallucination rate (2-sigma band) per model/task
import pandas as pd
df = pd.DataFrame(gens["train"])
headline = (
    df.groupby(["source_model", "task"])["factual_flag"]
      .mean()
      .rename("factual_hallucination_rate_2sigma")
      .reset_index()
)
print(headline)

Evaluation and repair code (BPHS scoring pipeline, Foldseek driver, Attention-Valley Resampling) is hosted separately at: https://anonymous.4open.science/r/plm-hallubench (MIT-licensed, executable, reviewer-accessible).

Intended use (construct validity)

PLM-HalluBench is built to measure cross-axis coherence of PLM generations — the disagreement between what a sequence looks like at the residue level and what it looks like when folded. The construct "biological plausibility under a natural reference distribution" is made concrete by:

• BPHS = (pLDDT + pTM)/2 − S_seq, thresholded at the µ_ref ± 2σ_ref band of the 5,000 natural references
• Faithful track: a generation passes only if at least one of its top-20 Foldseek neighbours carries the prompted InterPro / Pfam annotation

Construct validity is supported by an independent PyRosetta ref2015 full-atom energy correlation (Spearman ρ ≈ −0.40 stratified, −0.35 random holdout), documented in the paper's physics-validation appendix.

In-scope uses

1. Comparative benchmarking of PLM hallucination rates across architectures and generation regimes
2. Methodological studies of sequence–structure coherence metrics
3. Training-time or decoding-time interventions that use BPHS as an internal signal

Out-of-scope uses

1. Certifying that any individual generated sequence is safe, foldable, or functional in vitro
2. Ranking of natural proteins by biological validity
3. Direct selection of candidates for synthesis without downstream experimental validation

Limitations

• Pfam-centric conditions — globular families dominate; membrane proteins, IDRs, and multi-domain assemblies are not separately stratified
• ESMFold as structure proxy — topologies outside ESMFold's training distribution may be mis-flagged
• Rosetta ref2015 as energy proxy — classical approximation, not experimental stability
• Reference distribution from 5 homologs — conservative band but noisier for rare families
• Prompt coverage limited to four formats — models requiring structure tokens, MSAs, or other modalities are not separately accommodated

Biases

• Length bias: retrieval prefers shorter sequences (tractable ESMFold/Rosetta) → underrepresents >500-residue and multi-domain proteins
• Taxonomy bias: overrepresents model organisms (H. sapiens, E. coli, M. musculus, S. cerevisiae)
• Function-class bias: dense Swiss-Prot families (kinases, EGF-like domains, Ig folds) better populated than niche families
• Annotation-transfer bias: inherits upstream Pfam/InterPro/Swiss-Prot annotation errors
• Generator bias: scores reflect the six specific PLMs at fixed sampling settings, not all plausible PLM outputs

Personal / sensitive information

None. Only protein sequences, structural confidences, and family-level annotations from public, reviewed biological databases. No human subjects, no identifying metadata.

Synthetic data

Yes (75 000 / 80 000 records). Generated sequences are clearly labelled source = "generated" and carry source_model, task, and seed fields. Natural references are labelled source = "swissprot". The two pools are never mixed during scoring: natural defines the reference band, synthetic is the object of evaluation.

Provenance

Derived from

• UniProtKB / Swiss-Prot — homolog retrieval, reference pool
• InterPro — functional annotations, keyword links
• Pfam — family identifiers, hierarchy
• AlphaFold / Swiss-Prot — structural-neighbour pool (via Foldseek)

Pipeline

1. Pfam families → InterPro keyword expansion → UniProt/Swiss-Prot homolog retrieval (5 per condition, shorter-sequence preference)
2. GPT-5 deterministic rendering of four prompt formats (no scoring role)
3. 5 generations per (model, task, condition) triple at fixed sampling settings
4. ESMFold pLDDT/pTM + 8-component S_seq → BPHS → µ_ref ± 2σ_ref threshold
5. Foldseek top-20 against AlphaFold/Swiss-Prot for the Faithful track
6. PyRosetta ref2015 full-atom energy on a stratified sample for construct validation

No crowdsourced annotation, no human labelling, no human subjects at any stage.

Social impact

Positive — reliable multi-level hallucination detection lowers the risk that biologically invalid PLM outputs advance into downstream design pipelines (drug design, enzyme engineering, therapeutics), and gives a shared evaluation substrate independent of any single generator.

Potential negative — cross-axis coherence signals could in principle reward generation of proteins with harmful function (toxins, biothreats). This release is scoped as evaluation + detection/repair; no new generative model or weights are released. The repair procedure (AVR) operates on an existing frozen PLM without expanding generative capability. Downstream wet-lab use should follow the biosafety review channels that already apply to protein design research.

Citation

@inproceedings{plmhallubench2026,
  title     = {PLM-HalluBench: A Multi-level Benchmark for Evaluating Hallucinations in Protein Language Models},
  author    = {Anonymous Authors},
  booktitle = {NeurIPS 2026 Evaluations and Datasets Track},
  year      = {2026}
}

License

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. Upstream resources retain their own licenses (UniProt: CC BY 4.0; InterPro: CC0 1.0; Pfam: CC0 1.0; AlphaFold / Swiss-Prot: CC BY 4.0; ESMFold / Foldseek / PyRosetta are not redistributed — users must install them under the developers' terms).

Maintenance

• Versioning: datasets are pinned by the Git tag / HF revision; all BPHS thresholds are recomputed whenever the natural reference pool is updated.
• Errata: submit issues at the anonymous code repository during the review period; a public tracker will replace the anonymous one after acceptance.
• Long-term hosting: Hugging Face is the canonical mirror; the dataset is fully self-contained (no external API calls) except for downstream Foldseek and ESMFold runs which users perform themselves.