ai21labs/LIFTBench

LIFT-Bench

Localized Instruction-Following Tasks (LIFT-Bench) is a benchmark of synthetic document-drafting tasks. A task gives the model a topic and an outline, a nested tree of headers whose leaves hold paragraphs, code blocks, tables and lists, and attaches constraints to individual elements of it.

Every constraint is scoped: it is scored only against the element it is attached to. That distinguishes "a paragraph in French somewhere" from "section 2.1 in French", and makes constraint density a generation parameter rather than an afterthought.

Each task mixes verifiable checks, run in code, with judged content requirements, run by an LLM judge, in a single weighted score. A response that meets every formatting constraint but skips the content still fails every judged constraint.

The scoring code, the verifiers taskset that runs it and the pipeline that synthesized the data are at github.com/AI21Labs/lift-bench.

Layout

The dataset has two axes. Each config is one of the ten generated datasets, and each config has a train and a test split. The paper's LIFT-Bench is the test split of every config (784 tasks), and LIFT-Train is the train split of every config (34,258 tasks).

from datasets import load_dataset

bench = load_dataset("ai21labs/LIFTBench", "short_doc_few_constraints", split="test")
train = load_dataset("ai21labs/LIFTBench", "short_doc_few_constraints", split="train")

The configs come from three generation runs: short outlines, long outlines and flat (no outline). Within a run every config shares the same topics and outlines; only the number of constraints attached to them changes.

Config	dataset_id	Train	Test
short_doc_few_constraints	l3_d2_sh3_sv2_snv2	3,681	85
short_doc_some_constraints	l3_d2_sh5_sv3_snv2	3,681	85
short_doc_many_constraints	l3_d2_sh7_sv3_snv3	3,681	85
short_doc_numerous_constraints	l3_d2_sh7_sv4_snv4	3,681	85
long_doc_few_constraints	l5_d2_sh5_sv2_snv2	2,978	66
long_doc_some_constraints	l5_d2_sh7_sv3_snv2	2,978	66
long_doc_many_constraints	l5_d2_sh7_sv3_snv3	2,978	66
long_doc_numerous_constraints	l5_d2_sh7_sv4_snv4	2,978	66
flat_doc_some_constraints	flat_sv3_snv3	3,811	90
flat_doc_many_constraints	flat_sv5_snv5	3,811	90

dataset_id encodes the generation parameters:

Code	Meaning
l<n>, d<n>	Outline shape: up to n children per header, n levels deep
sh<n>	Up to n sections carry constraints
sv<n>, snv<n>	Up to n verifiable and n judged constraints per section (per task, for flat)
flat	No outline; every constraint applies to the whole response

Columns

Column	Type	Content
prompt	list of messages	The task shown to the model, one user message
info	dict	{"evaluations": [...]}, the rubric the task is scored against
example_id	string	Identifier of the task
topic	string	Seed topic, a Wikipedia article title
document_outline	string	The markdown outline the task was built from
dataset_name	string	The config name
dataset_id	string	The generation-parameter code
split	string	train or test
metadata	dict	Empty

Each entry of info.evaluations is one constraint:

{
  "evaluation_id": "content_verifiable_constraints_3",
  "weight_group": "content",
  "weight": 0.175,
  "metric_name": "lexical_count",
  "metric_args": { "string": "climate", "lower_bound": 2, "upper_bound": 5 },
  "parser_name": "nested_html",
  "parser_args": { "selector_lst": [["H2", 1]], "is_shallow": false, "document_type": "markdown" }
}

parser_name and parser_args name the section of the document the constraint is scored on; metric_name and metric_args name the check. The score of a task is the weighted sum of its per-constraint scores.

Scoring

Score a model with the drafting_constraint taskset from the repository, which runs under verifiers:

hf download ai21labs/LIFTBench --repo-type dataset --local-dir data/splits
ENV_ID=drafting_constraint VLLM_MODEL=<model-name> ./run_eval.sh \
    --env.taskset.dataset-path data/splits/short_doc_few_constraints/test.jsonl

Judged constraints are scored by the LLM judge each task names in metric_args.critic_model_name, google/gemma-4-31B-it, called through an OpenAI-compatible endpoint. --env.taskset.judge-model <model-name> scores them with another judge instead, without changing the data. Results with a different judge are not comparable to the published ones.

Loading notes

• The rubric fields metric_args and parser_args.selector_lst hold values whose type varies between rows, which datasets stores as JSON fields and decodes on access. Loading was verified with datasets 4.8 and 5.1.
• When datasets loads the files it re-encodes them, and constraint weights come back rounded to 10 decimal places (0.16666666666666666 becomes 0.1666666667). A task's score changes by less than 1e-9 as a result. The JSONL files themselves hold the exact values, and the verifiers taskset reads them directly.

License and attribution

The data is released under the Apache 2.0 license. Topics are titles of Wikipedia's vital articles, available under CC BY-SA 4.0.

Citation