Hauser7733/TAPE_Eval_v1

TAPE Eval v1 — Polish v2 MCQ + Ranking Bundle

Unified evaluation bundle for TAPE (Rao et al. 2019) zero-shot chat-LLM evaluation (sieval Polish v2 Type 3 task redesign). Consolidates four MCQ configs and four Ranking configs spanning three TAPE sub-tasks.

Canonical consumer: sieval.datasets.TAPEEvalDataset(subset=<config>).

Configs

12 configs total across three MCQ flavours plus N-way ranking:

• Cutoff MCQ (2-way): single-sample absolute-threshold classification
• Find-the-X MCQ (N-way): comparative discrimination among N candidates (1 bright + N-1 dark etc.), balanced 50/50 between "which is bright" / "which is dark" per config
• Rank (N-way): windowed stratified sampling, rank by fitness high-to-low

Config	Format	N options	Source test split	N items
fluorescence_mcq	MCQ (cutoff)	2	AI4Protein/TAPE_Fluorescence (27,217 → 500 class-balanced)	500
fluorescence_mcq4	MCQ (find)	4	AI4Protein/TAPE_Fluorescence (balanced qtypes)	500
fluorescence_mcq10	MCQ (find)	10	AI4Protein/TAPE_Fluorescence (balanced qtypes)	500
stability_mcq	MCQ (cutoff)	2	AI4Protein/TAPE_Stability (12,851 → 500 class-balanced)	500
stability_mcq4	MCQ (find)	4	AI4Protein/TAPE_Stability (balanced qtypes)	500
stability_mcq10	MCQ (find)	10	AI4Protein/TAPE_Stability (balanced qtypes)	500
remote_homology_mcq4	MCQ	4	Hauser7733/TAPE_Remote_Homology (3 holdouts full)	3,244
remote_homology_mcq10	MCQ	10	Hauser7733/TAPE_Remote_Homology (3 holdouts full)	3,244
fluorescence_rank4	Rank	4	AI4Protein/TAPE_Fluorescence (windowed stratified)	500
fluorescence_rank10	Rank	10	AI4Protein/TAPE_Fluorescence (windowed stratified)	500
stability_rank4	Rank	4	AI4Protein/TAPE_Stability (windowed stratified)	500
stability_rank10	Rank	10	AI4Protein/TAPE_Stability (windowed stratified)	500

Cutoffs (MCQ — binary classification)

• Fluorescence: log-fluo = 2.5 — Shanehsazzadeh et al. 2020 (arxiv:2011.03443)
• Stability: stability score = 1.0 — Rocklin et al. 2017 (Science 357:168): "stability score > 1.0 = successfully folded"

Ranking — Windowed Stratified Sampling

Tier ratio hard : medium : easy = 1 : 2 : 1 (125 / 250 / 125 per config).

• window_frac(rank4) = {hard: 0.15, medium: 0.30, easy: 0.60}
• window_frac(rank10) = {hard: 0.25, medium: 0.50, easy: 1.00}
• min_gap = 0.5 * assay_std / N (prevents ties)
• Fallback to quantile-anchored jitter after 3 window-expansion retries

Windowed stratified design draws from FaceNet semi-hard negative mining (Schroff et al. 2015) and Fogliato et al. 2024 stratification (arxiv:2406.07320). LLM answer format is comma-separated candidate ids wrapped in [ANSWER]...[/ANSWER] (RankGPT style, arxiv:2304.09542).

Remote Homology — v0.2 SCOPe hierarchy integration

Upstream data exposes only integer fold_label (0..1195). v0.2 attaches each fold's SCOPe 1.75 hierarchy id (e.g. a.1, c.1) and human-readable structural description (e.g. TIM beta/alpha-barrel) so that chat LLMs can use pretrained structural knowledge rather than memorise benchmark-internal integers. Option text format: {fold_id} — {description}.

Mapping sources (shipped verbatim in reference/):

• fold_label → fold_id: DeepSF fold_label_relation2.txt (Hou et al. 2018)
• fold_id → description: SCOPe 1.75 dir.des.scop.1.75.txt (Murzin/Brenner/Fox/Chandonia)
• Joined table: reference/scope_fold_descriptions.tsv (1196 rows)

Distractors are sampled uniformly from other fold_labels in the same evaluation split; TAPE RH test splits are heavily dominated by classes a (all-α) and b (all-β), so distractors inherently share SCOPe class with the golden fold in most samples. Design doc: PROMPT_DESIGN_TAPE_MCQ.md §3.2.2 v0.2 section in the sieval worktree.

Schema

All configs share the universal fields (sample_id, task, eval_format, n_options); MCQ configs add options_json / answer / label; Rank configs add candidates_json / golden_ranking_json / tier / avg_adjacent_gap / fallback. Task-specific context (wt_seq, topology, protein_length, fold_id, etc.) is included as applicable.

fitness_hidden values embedded in Rank candidates_json are the ground-truth continuous scores and must not be surfaced to the LLM — the sieval loader enforces this via the task's preprocess() method.

Usage

from datasets import load_dataset

ds = load_dataset("Hauser7733/TAPE_Eval_v1", "fluorescence_mcq", split="test")

# In sieval:
from sieval.datasets import TAPEEvalDataset
ds = TAPEEvalDataset(subset="stability_rank10")

Build tooling + reference data (in-repo)

For reproducibility, this repo ships its own build and upload tooling under scripts/, plus the reference data files used during construction under reference/:

• scripts/build_tape_mcq.py — main dataset builder (sieval-unaware; drives the four sub-commands fluorescence / stability / remote_homology / fluorescence_nway / stability_nway / fluorescence_rank / stability_rank)
• scripts/upload_hf.py — this file, the HF publisher
• scripts/make_showcase.py — reads sieval TaskRunner shards and emits a sample_showcase.json for qualitative audit
• reference/cutoffs.md — literature sources for Fluo/Stab cutoffs
• reference/stab_topology_filter.txt — the designed-topology allow-list used to guard against stability memorization shortcuts
• reference/sampling_config.yaml — global seeds / ratios / window fractions
• reference/scope_fold_descriptions.tsv — 1196-row joined label → fold_id → description table (v0.2 RH upgrade)

To rebuild from scratch:

python scripts/build_tape_mcq.py fluorescence --n-samples 500 --out data/
python scripts/build_tape_mcq.py stability --n-samples 500 --out data/
python scripts/build_tape_mcq.py fluorescence_nway --n-options 4 --out data/
python scripts/build_tape_mcq.py fluorescence_nway --n-options 10 --out data/
python scripts/build_tape_mcq.py stability_nway --n-options 4 --out data/
python scripts/build_tape_mcq.py stability_nway --n-options 10 --out data/
python scripts/build_tape_mcq.py remote_homology --n-options 4 --out data/
python scripts/build_tape_mcq.py remote_homology --n-options 10 --out data/
python scripts/build_tape_mcq.py fluorescence_rank --n-options 4 --out data/
python scripts/build_tape_mcq.py fluorescence_rank --n-options 10 --out data/
python scripts/build_tape_mcq.py stability_rank --n-options 4 --out data/
python scripts/build_tape_mcq.py stability_rank --n-options 10 --out data/
python scripts/upload_hf.py

Design rationale lives in the sieval Polish v2 worktree's PROMPT_DESIGN_TAPE_MCQ.md.

Citations

• Rao et al. 2019. Evaluating Protein Transfer Learning with TAPE. NeurIPS.
• Shanehsazzadeh et al. 2020. Is Transfer Learning Necessary for Protein Landscape Prediction?
• Rocklin et al. 2017. Global analysis of protein folding.... Science 357:168.
• Schroff et al. 2015. FaceNet: A Unified Embedding.... arxiv:1503.03832.
• Sun et al. 2023. RankGPT. EMNLP. arxiv:2304.09542.