BioSum-CUH

A Biography Summarization Benchmark with Token-Level Correctness, Uncertainty, and Hallucination Annotations

BioSum-CUH is a benchmark for studying factual generation over biography contexts. It combines biography-based question answering and structured summarization with token-aligned model predictions, final-layer attention activations, logit statistics, and three-way factuality annotations.

The release supports two complementary research settings:

1. Biography understanding: read a biography and recover one or more target attributes, including dates, places, education, employment, and other personal-profile fields.
2. Token-level reliability modeling: predict whether each generated token is correct/functional, hallucinated or contradictory, or part of an abstaining/unknown response.

The benchmark data and activation data are related but are not one-to-one. The benchmark contains 10,000 training examples, whereas the activation training split contains 7,767 generation traces derived from 5,500 unique training examples. The separate detector_eval split contains 139 traces derived from 103 benchmark-validation examples. Repeated traces are expected because some source examples were collected under multiple generation or context-selection settings.

Dataset configurations

Configuration	Splits	Purpose
benchmark	train, validation, test	Biography contexts, questions, and reference answers
activations	train, detector_eval	Full token-level annotations, predictions, logits, and attention activations
activation-metadata	train, detector_eval	Lightweight projection for Dataset Viewer inspection and metadata analysis

benchmark is the default configuration. The full activation configuration is large because it preserves float32 tensors. Use streaming when only a subset is needed.

Benchmark splits

Split	Examples	Original subset	Task design
train	10,000	data_analysis	Short synthetic biographies with one queried attribute
validation	120	ed1/data_analysis	Six attribute types, with 20 examples per type
test	100	data_long	Long-context, six-attribute biography summarization

The test split in this repository is the selected 100-example data_long split. It should not be confused with the separate 30-example data_long_long_common_short subset used by some internal evaluation runs.

Character-length statistics for the released biography text are shown below. These are Unicode character counts, not tokenizer-specific token counts.

Split	Minimum	Median	Mean	Maximum
train	986	1,017	1,020.2	1,157
validation	994	1,012	1,010.9	1,030
test	10,078	40,317	46,134.3	98,195

The training split covers 22 single-attribute question types:

marry date, job title, current city, email address, phone number, company, favorite color, user agent, credit card provider, currency used, catch phrase, street address, vehicle license plate, favorite file extension, domain name, cryptocurrency, timezone, isbn code, lucky number, hobby, marital status, and personality type.

The validation and test data focus on six biography attributes:

1. birth date;
2. birth place;
3. university;
4. major;
5. company;
6. work place.

Benchmark schema

Each row in the benchmark configuration contains:

Field	Type	Description
example_id	int64	Stable row identifier within the split
biography	string	Biography context presented to the model
name	string	Target person or entity name
question	string	Requested attribute or ordered attribute list
answer	string	Reference answer
prompt_template	string, nullable	Original prompt template when present in the source split

The test answers use an ordered six-field format. Models should preserve the requested order so that partial correctness can be evaluated per attribute.

Example:

{
  "example_id": 0,
  "name": "Emma Thompson",
  "question": "(1) birth date, (2) birth place, (3) university, (4) major, (5) company, (6) work place",
  "answer": "(1) 12/29/1880, (2) Tokyo, (3) Harvard University, (4) Computer Science, (5) Google, (6) San Francisco Bay Area",
  "biography": "...",
  "prompt_template": "Read the biography, and answer the question one by one..."
}

Activation and token-annotation data

The activations configuration contains 7,767 aligned training records and 139 held-out detector-evaluation records. The records contain predictions produced with Meta-Llama-3.1-8B-Instruct and the corresponding layer-31 attention outputs. All tensor columns are padded to 30 generated-token positions; length marks the valid prefix.

Detector evaluation split

detector_eval is the held-out token-classification evaluation split used by the accompanying uncertainty-detector experiment. It is derived from the BioSum-CUH benchmark/validation examples, not from RULER or another external benchmark, and is distinct from the 100-example long-context benchmark/test split used for final QA and end-to-end evaluation.

The release preserves all 139 collected trajectories over 103 unique validation examples, totaling 1,631 valid output tokens. The paper protocol excludes the first 22 free-form trajectories and evaluates activation_id >= 22, which leaves 117 trajectories over 88 unique validation examples and 1,454 valid tokens. Keeping the excluded rows in the release makes the original filtering explicit and auditable.

At the exact (name, question, answer) level, the 103 unique source examples overlap with 103 rows in benchmark/validation and with zero rows in either benchmark/train or benchmark/test.

Scope	Trajectories	Unique source examples	Tokens	Label 0	Label 1	Label 2
Complete detector_eval split	139	103	1,631	400	1,052	179
Paper protocol (activation_id >= 22)	117	88	1,454	400	1,036	18

Two legacy source annotation rows, activation IDs 55 and 58, each contained one redundant trailing class-1 label beyond the recorded prediction and length. The Parquet conversion removes only those two out-of-range labels, consistent with the length mask used by the original evaluator.

Backbone and model specificity

The released model-dependent artifacts were collected from Meta-Llama-3.1-8B-Instruct (Llama-3.1-8B-IT):

• transformer layer: L31, the final layer in the 32-layer model;
• attention-output width: 4096;
• activation field: attn_act, with padded shape [30, 4096] per record;
• logit fields: top-1-minus-top-2 margin and descending top-50 raw logits;
• token alignment: decoded pieces from the Llama 3 Instruct tokenizer.

These artifacts must not be treated as interchangeable features for another model. To use BioSum-CUH with a different model family, model size, checkpoint revision, layer, or tokenizer, reuse the benchmark biographies/questions and reference answers, but reconstruct all model-dependent records. This includes:

1. generating new predictions with the target model;
2. recording the target model's token-aligned activations at a documented layer;
3. recording its logits and recomputing logit_margin and logit_topk;
4. recomputing length and prediction_tokens with the target tokenizer;
5. assigning new token-level labels to the new generated token sequence; and
6. validating alignment and publishing a separate model-specific configuration with its model identifier, revision, tokenizer, layer, shape, and dtype.

Zero-padding or truncating another model's hidden states to 4096 dimensions does not reproduce the released Llama activation distribution and should not be presented as an equivalent reconstruction.

Field	Hugging Face feature	Description
activation_id	Value("int64")	Stable record ID within the activation split: 0--7,766 for train and 0--138 for detector_eval
source_train_index	Value("int64")	Legacy field name: matching benchmark train row for activation train, or benchmark validation row for detector_eval
source_occurrence	Value("int8")	Zero-based occurrence number for repeated collection of the same source row
collection_method	Value("string"), nullable	Collection mode, included only when recoverable from provenance
name	Value("string")	Target name
question	Value("string")	Queried attribute
answer	Value("string")	Reference answer
prediction_tokens	List(Value("string"))	Tokenizer-decoded generated pieces in order
labels	List(Value("int8"))	Unpadded token labels, aligned one-to-one with prediction_tokens
length	Value("int32")	Number of valid generated-token positions, from 8 to 30
attn_act	Array2D((30, 4096), "float32")	Padded layer-31 attention-output activation
logit_margin	Array2D((30, 1), "float32")	Padded top-1 minus top-2 logit margin
logit_topk	Array2D((30, 50), "float32")	Padded descending top-50 raw logits

prediction_tokens stores decoded token pieces rather than vocabulary IDs. Leading spaces and punctuation are meaningful and should be preserved.

activation-metadata contains the same IDs, source mappings, predictions, labels, lengths, and logit metadata, but omits attn_act. It is intended for quick inspection rather than as a separate source of truth.

Token-label semantics

Label	Name	Meaning
0	hallucinated / incorrect	Wrong, contradictory, unsupported, or predicate-mismatched asserted content
1	correct / functional	Correct answer or name content, plus punctuation, spacing, and functional tokens that do not assert an incorrect fact
2	unknown / abstaining	Tokens expressing lack of knowledge, missing information, or a request for additional information

The released detector task preserves and predicts all three classes. For the paper's aggregate uncertainty recall, labels 0 and 2 are grouped as uncertain, while label 1 is treated as certain; this reporting-time aggregation does not turn the main detector into a binary classifier.

These labels describe the semantic status of generated text. They are not direct measurements of calibrated epistemic or aleatoric uncertainty.

Label statistics

There are 122,453 valid labeled tokens:

Label	Tokens	Share
0	19,832	16.196%
1	101,924	83.235%
2	697	0.569%

The substantial class imbalance, especially the small number of label-2 tokens, should be considered when choosing metrics and sampling strategies.

Annotation and processing pipeline

1. Biography question-answer examples were constructed from synthetic target profiles. Training profiles use multilingual Faker locales to diversify names and attribute values, while prompts and questions are in English.
2. Reader-model outputs were collected under multiple generation/context settings. For each generated token, the pipeline recorded the decoded token, final-layer attention output, logits, and source-example metadata.
3. A GPT-OSS-120B-assisted labeling prompt assigned token-level labels according to the three-way rules above.
4. Records were padded to 30 token positions. Labels and predictions remain unpadded lists in the Hugging Face representation, and length preserves the valid tensor boundary.
5. The aligned artifacts were converted from PyTorch .pth/JSONL files into schema-defined, sharded Parquet files.

The following integrity checks were applied to the source artifacts:

• 7,767 records in every aligned artifact;
• exact len(labels) == len(prediction_tokens) == length for every record;
• tensor shapes of [30, 4096], [30, 1], and [30, 50];
• all 7,767 activation rows matched exactly one training row by (name, question, answer);
• 5,500 unique training rows represented, with at most three activation occurrences per source row;
• logit_margin == logit_topk[..., 0] - logit_topk[..., 1] bit-for-bit over all valid positions;
• zero padding after length in the source tensor artifacts.

Loading the data

Install the datasets package and load the benchmark configuration:

from datasets import load_dataset

dataset = load_dataset("Saria307/biosum-cuh", "benchmark")

train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

print(train[0]["question"])
print(train[0]["answer"])

Stream the full activation configuration to avoid downloading the complete activation table before reading the first record:

from datasets import load_dataset
import numpy as np

stream = load_dataset(
    "Saria307/biosum-cuh",
    "activations",
    split="train",
    streaming=True,
)

row = next(iter(stream))
length = row["length"]

attn_act = np.asarray(row["attn_act"], dtype=np.float32)[:length]
logit_margin = np.asarray(row["logit_margin"], dtype=np.float32)[:length]
logit_topk = np.asarray(row["logit_topk"], dtype=np.float32)[:length]
labels = np.asarray(row["labels"], dtype=np.int64)

assert attn_act.shape == (length, 4096)
assert logit_margin.shape == (length, 1)
assert logit_topk.shape == (length, 50)
assert labels.shape == (length,)

Load the held-out records and reproduce the exact paper evaluation scope:

from datasets import load_dataset

detector_eval = load_dataset(
    "Saria307/biosum-cuh",
    "activations",
    split="detector_eval",
)
paper_eval = detector_eval.filter(lambda row: row["activation_id"] >= 22)

assert detector_eval.num_rows == 139
assert paper_eval.num_rows == 117
assert sum(paper_eval["length"]) == 1454

For inspection without downloading the 4096-wide activation tensors, use the matching lightweight split:

detector_eval_metadata = load_dataset(
    "Saria307/biosum-cuh",
    "activation-metadata",
    split="detector_eval",
)

To construct a new train/validation partition from the activation training records, split by source_train_index, not by activation_id. Otherwise repeated collections of the same source example can leak across partitions. Do not combine detector_eval with the training records during fitting, checkpoint selection, threshold selection, or hyperparameter tuning.

Recommended evaluation

Biography task

For single-attribute examples, report normalized exact match and token-level F1. For the six-attribute test examples, also report per-attribute accuracy and the macro average across the six requested fields. Semantic variants of dates, locations, and organization names may require careful normalization or a documented judge model.

If an LLM judge is used, report the judge model and revision, full prompt, temperature, parsing/failure policy, and whether missing outputs receive zero. Judge-based scores should not be compared unless these settings are held fixed.

Token-level reliability task

For three-way prediction, report per-class precision, recall, F1, macro F1, and the confusion matrix. For binary uncertain-vs-certain prediction, explicitly state the label mapping; the accompanying workflow uses {0, 2} -> uncertain and {1} -> certain. Because the classes are imbalanced, accuracy alone is not sufficient. When comparing with the paper detector results, use the detector_eval split with activation_id >= 22 and report that filter.

Intended uses

BioSum-CUH is intended for research on:

• long-context biography summarization and attribute extraction;
• factuality and hallucination detection during generation;
• token-level uncertainty and abstention modeling;
• logit-based confidence baselines;
• activation-based probing and lightweight detector training;
• context-compression and KV-cache methods evaluated under a shared benchmark;
• analysis of correct, incorrect, and unknown generation behavior.

Out-of-scope uses

The dataset is not intended for:

• making claims about real people;
• identity verification, background checks, surveillance, or profiling;
• making medical, legal, financial, employment, credit, or other high-impact decisions;
• treating machine-generated labels as definitive human factuality judgments;
• comparing model uncertainty across tokenizers or architectures without accounting for representation and calibration differences.

Limitations

• Synthetic targets: Target profiles and many attributes are synthetic. Synthetic names and contact-like values can coincidentally resemble real people or real identifiers and must not be interpreted as verified facts.
• Long-context construction: Test contexts may include repeated filler and Wikipedia-derived distractor passages. Success may therefore reflect robustness to redundancy and distractors as well as biography understanding.
• Distribution shift: Training and validation biographies are approximately 1,000 characters, while test biographies are substantially longer. The split design intentionally contains a strong length and task-format shift.
• Machine-assisted labels: Token labels follow a particular prompt and labeling policy. The policy gives lexical matches and functional tokens priority, so token labels need not coincide with a sentence-level factuality judgment.
• Class imbalance: Unknown/abstaining tokens are rare.
• Repeated activation sources: The activation set contains repeated collections of some training examples. Random row-level splits can leak source information.
• Incomplete collection provenance: The merged activation artifacts do not reliably expose every original collection mode. Missing modes are left null rather than inferred from row order.
• Detector-evaluation filtering: The paper detector protocol excludes activation IDs 0--21 from detector_eval. They remain publicly available, and users should state whether they evaluate all 139 trajectories or the 117-trajectory paper subset.
• Model specificity: The released activations and logits are specific to Meta-Llama-3.1-8B-Instruct, its tokenizer, final layer L31, and the collection pipeline. They should not be treated as model-independent features. Using another model requires reconstructing predictions, token labels, activations, logits, and lengths from that model.
• Fixed output window: Tensor features preserve at most 30 generated-token positions per record.
• Judge sensitivity: Semantic evaluation can vary with answer normalization or judge-model choice.

Ethical and privacy considerations

The benchmark is designed for controlled research rather than representing real biographies. Nevertheless, generated names, addresses, phone-like strings, and other profile fields can accidentally overlap with real-world data. Users should avoid attempts to identify individuals and should not use the dataset for decisions about people.

Long test contexts may contain Wikipedia-derived passages about real people. Those passages can contain the biases, omissions, inaccuracies, or outdated information present in their source material. Source URLs should be retained for attribution and auditability.

Licensing and attribution

This is a mixed-provenance release:

• project-authored benchmark packaging, metadata, annotations, and conversion code are released under the Apache License 2.0;
• Wikipedia-derived text remains subject to the Creative Commons Attribution-ShareAlike license and any applicable source terms;
• generated predictions and model-derived activations may also be subject to the terms of the model used to produce them, including the Llama 3.1 license.

For this reason, the Dataset Card uses license: other rather than applying Apache-2.0 to every text span in the repository. Users are responsible for preserving attribution and complying with all applicable upstream terms. See DATA_LICENSE.md for the release-level notice.

Citation

If you use BioSum-CUH, cite the dataset repository:

@misc{biosum_cuh_2026,
  title        = {BioSum-CUH: A Biography Summarization Benchmark with Token-Level Correctness, Uncertainty, and Hallucination Annotations},
  author       = {{BioSum-CUH Contributors}},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/Saria307/biosum-cuh}
}

When a corresponding paper or archival release becomes available, add its citation alongside the repository citation above.

Contact

For questions, corrections, or removal requests, open a discussion at:

https://huggingface.co/datasets/Saria307/biosum-cuh/discussions