romerocruzsa's picture
docs/fix: Typo on failed models for "table and unit reasoning"
9671191 verified
|
Raw
History Blame Contribute Delete
8.13 kB
metadata
license: other
language:
  - en
pretty_name: BioManufacturingBench v1.0.0
size_categories:
  - 1K<n<10K
task_categories:
  - question-answering
  - visual-question-answering
  - text-generation
tags:
  - biomanufacturing
  - biotechnology
  - multimodal
  - benchmark
  - objective-scoring
  - temporal-holdout
configs:
  - config_name: default
    data_files:
      - split: test
        path: data.jsonl
Capicú Technologies

BioManufacturingBench v1.0.0

BioManufacturingBench v1.0.0 is a 2,000-item benchmark for evidence-grounded biomanufacturing reasoning. It covers evidence extraction, mass-balance calculation, process diagnosis, microscopy count-range estimation, strict output formatting, and abstention. Every primary score is computed by a deterministic rule; no score uses an LLM judge. Public records are deliberately answer-free so the benchmark remains useful for future evaluation.

Usage

from datasets import load_dataset

bench = load_dataset("capicu-ai/BioManufacturingBench", split="test")

Records with input_type="image_text" are the 400 microscopy questions. Text-only systems should be evaluated on the remaining 1,600 records and reported separately. Assigned-task scores are not directly comparable when one model receives images and another does not.

Composition

Task family Items Evidence source Primary scorer
Evidence extraction 400 Held-out 2026 articles Semantic JSON, candidate selection, classification accuracy
Table and unit reasoning 350 Deterministic mass balances Numerical tolerance
Process trajectory reasoning 500 Deterministic simulation Closed-label exact match and semantic JSON
Microscopy reasoning 400 Reserved microscopy groups Exact count-range label
Instruction following 250 Reserved source metadata Exact JSON structure
Abstention 100 Balanced answerable and incomplete records Exact match

500 fresh held-out open-access articles from 2026 supply the text evidence. 3,927 reserved microscopy images are indexed for the image items. Source groups used here are excluded from the BioManufacturingCorpus and BioManufacturingInstruct training releases, so a model trained on those releases has not seen this evidence through that path.

Construction

Each family is built from a frozen deterministic recipe rather than free-form question writing. Answers derive from exact evidence, closed-form calculations, deterministic trajectory diagnostics, image-mask count ranges, or strict JSON rules. Process questions use explicit answer vocabularies, microscopy uses broad count bins rather than unreasonable exact high-count matching, and abstention records are balanced between answerable and incomplete cases. Public prompts and private answers remain in separate governed locations.

Evaluation protocol

Use deterministic decoding where the provider permits it. Record the model revision, provider, prompt version, generation settings, failure count, and date. Treat failed requests as missing observations and report them separately rather than counting infrastructure failures as model errors.

Report at minimum: overall score, text-only score, per-task-family scores, multimodal coverage and microscopy score, request completion rate, and uncertainty intervals for any full study.

Preliminary baseline

A stratified 100-item sample was run against 16 open-weight checkpoints. It contains 80 text items and 20 microscopy items. Text-only models received 80 requests and multimodal models received 100, for 1,440 total requests. All 1,440 returned scored answers.

The pilot is evidence of benchmark behavior, not a claim that 100 sampled items establish a definitive model ranking. Scores below are deterministic and use each model's assigned tasks. Microscopy is shown separately because text-only models did not receive those items.

Model Assigned-task score Evidence Table/unit Process Microscopy Instruction Abstention
Kimi K3 96.0% 95.0% 100.0% 100.0% 85.0% 100.0% 100.0%
GPT-OSS 120B 87.5% 95.0% 72.2% 86.4% 90.0% 100.0%
Inkling 82.0% 90.0% 55.6% 90.9% 75.0% 90.0% 100.0%
MiniMax M3 78.9% 94.7% 27.8% 86.4% 85.0% 90.0% 100.0%
DeepSeek V4 Pro 59.8% 79.4% 5.6% 77.3% 70.0% 70.0%
Qwen 3.5 397B 59.8% 78.8% 0.0% 86.4% 40.0% 70.0% 100.0%
GPT-OSS 20B 59.2% 71.7% 16.7% 63.6% 70.0% 90.0%
Qwen 3.6 27B 58.9% 74.7% 0.0% 68.2% 60.0% 70.0% 100.0%
DeepSeek V4 Flash 57.3% 79.1% 5.6% 63.6% 80.0% 70.0%
Qwen 3.6 35B-A3B 56.6% 77.8% 0.0% 72.7% 55.0% 70.0% 70.0%
Qwen3 235B-A22B 55.2% 70.9% 5.6% 54.5% 70.0% 100.0%
GLM-5.2 55.0% 40.0% 5.6% 81.8% 70.0% 100.0%
Llama 3.3 70B 54.3% 82.2% 0.0% 50.0% 70.0% 90.0%
Laguna-S-2.1 50.5% 62.2% 5.6% 50.0% 70.0% 90.0%
Qwen 3.5 9B 47.8% 73.8% 0.0% 68.2% 25.0% 50.0% 80.0%
Gemma 4 31B 43.0% 5.0% 0.0% 63.6% 45.0% 90.0% 100.0%

Task-family separation

Performance differed sharply by capability. Kimi K3 led the assigned-task pilot at 96.0% and was the only model to score 100% on evidence, table/unit, process, instruction, and abstention simultaneously. Table/unit reasoning remained highly selective: six models scored 0%, while Kimi K3 scored 100%, GPT-OSS 120B scored 72.2%, and Inkling scored 55.6%. Process reasoning ranged from 50.0% to 100.0%. Microscopy count-range accuracy ranged from 25.0% to 85.0% across the eight multimodal models.

Instruction-following and abstention are reported as competency checks. They are useful for detecting failures but are not intended to dominate model ranking.

Fields

Field Description
id Stable public question identifier
task One of the six task families
prompt Text shown to the model
input_type text or image_text
image Governed image URI for microscopy items; empty for text items
scorer Frozen objective scoring rule
source_id Source ancestry identifier
source_url Public source location where applicable
license Record-level source license where applicable
evaluation_role discrimination or competency
release_version Always v1.0.0

The public file is data.jsonl. It contains no reference answers, internal build identifiers, or nullable scenario_id field.

Governance and license

This is a mixed-source benchmark, so the repository is released under license: other rather than asserting one blanket license across every source. Source-derived records retain their identifiers and URLs, and admission follows a fail-closed license policy. Users are responsible for the terms of the referenced source datasets.

Benchmark prompts, metadata, and project-authored deterministic scenarios may be used for research subject to the repository terms. The private answer key must not be redistributed or used for training.

Citation

Cite the dataset repository with the exact benchmark version and access date. A methods manuscript citation will replace this when released.

@misc{biomanufacturingbench2026,
  title  = {BioManufacturingBench v1.0.0},
  author = {S.A. Cruz Romero},
  year   = {2026},
  url    = {https://huggingface.co/datasets/capicu-ai/BioManufacturingBench},
  note   = {Version v1.0.0}
}