Extracted entities, classifications, answers, summaries, and embeddings from biomedical NLP.

Each job returns task-specific results — extracted entities, classifications, answers, summaries, or embeddings depending on what you ran. These turn unstructured text into structured knowledge you can feed into target identification, drug repurposing, or knowledge-graph construction.

Output artifacts

The transformers_inference tool wraps a generic Hugging Face pipeline() call and is not currently wired into the registered tool set or backed by a deployed Modal app — see the warning on Transformers on the platform. The field names below describe a standard Hugging Face pipeline's output shape for each task, not a verified platform-specific schema.

Each completed Transformers job returns task-specific results depending on which task you ran: ner, classification, zero_shot, qa, summarization, or embeddings. There is no separate relation-extraction task — infer relationships from co-occurring NER entities, or use qa against the source text.

NER output fields

Field	Type	Description
text	string	The original input text
entities	array	List of extracted entity objects
entity.word	string	The entity text span (e.g., "EGFR")
entity.entity_type	string	Entity category (Drug, Gene, Disease, Chemical, Mutation, Cell Line)
entity.start	integer	Character start position in the text
entity.end	integer	Character end position in the text
entity.score	float	Confidence score (0-1)
entity_counts	object	Frequency counts by entity type

Interpreting NER confidence scores

Score Range	Confidence	Interpretation
0.9-1.0	Very high	Entity almost certainly correct; standard biomedical terms
0.7-0.9	High	Entity likely correct; may need verification for ambiguous terms
0.5-0.7	Moderate	Entity possibly correct; check context for false positives
<0.5	Low	The pipeline still returns these; there is no confidence-threshold parameter to filter them

Common entity patterns

Multi-word entities: "non-small cell lung cancer" is a single Disease entity; check that boundaries are correct
Abbreviations: Both the full form and abbreviation may be extracted separately (e.g., "epidermal growth factor receptor (EGFR)" yields two Gene mentions)
Nested entities: Some entities may overlap (e.g., "EGFR T790M" contains both a Gene and a Mutation); both are reported
Negated entities: Entities in negated contexts ("not associated with BRCA1") are still extracted; check surrounding text for negation

Question answering and summarization output fields

Field	Type	Description
answer	string	Extracted answer span (`qa`) or generated summary text (`summarization`)
score	float	Confidence score (0-1); `qa` only
start / end	integer	Character offsets of the answer span in the source text (`qa` only)

Text classification output fields

Field	Type	Description
label	string	Predicted class label
score	float	Prediction confidence (0-1)
probabilities	object	Per-class probability distribution

Interpreting classification

High confidence (>0.8): Strong prediction; document clearly belongs to the assigned category
Moderate confidence (0.5-0.8): Document may have characteristics of multiple categories
Low confidence (~0.5): Near the decision boundary; manual review recommended

Embedding output fields

Field	Type	Description
embedding	array	Dense vector representation (768-dimensional for BERT models)
model	string	Model used for embedding generation
pool_strategy	string	Pooling method used

Using embeddings

Cosine similarity: Compare two document embeddings for semantic similarity
Clustering: Group documents by embedding proximity (k-means, HDBSCAN)
Semantic search: Find documents closest to a query embedding
Visualization: Project embeddings to 2D with t-SNE or UMAP for exploration

Batch output format

For multi-document batch jobs:

Each document gets its own entity list, classification, answer, summary, or embedding result
Results are keyed by document index or PMID
Aggregate statistics include total entity counts and most frequent entities

Downloading artifacts

JSON export: Structured annotations with all entity positions, types, and scores
CSV export: Entity table with columns for document, entity text, type, start, end, and score
Annotated HTML: Original text with color-coded entity highlighting
Embedding files: NPY or CSV format for downstream ML use

Canvas visualization

When a Transformers job completes, a result card appears on the canvas showing:

Annotated text with highlighted entities (color-coded by type)
Entity frequency bar chart
Classification result with confidence bar
Answer span or generated summary (for qa or summarization)

Quality checks

Verify that known entities in the input text are captured (recall check)
Check entity boundaries for multi-word terms (no partial words)
Confirm entity types are correct (gene vs. protein naming can be ambiguous)
Compare extracted entities across models (BioBERT vs. PubMedBERT) for consensus

Transformers outputs and interpretation