Datasets:

H2S-Research
/

Highlight-Then-Summarize

Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
schema_version: string
id: string
benchmark: string
ability: string
question: string
gt: string
input_token_num: int64
input_token_num_original: int64
input_token_num_mode: string
input: list<item: struct<role: string, content: string>>
  child 0, item: struct<role: string, content: string>
      child 0, role: string
      child 1, content: string
longbenchv2_reconstruction: struct<source_file: string, source_data_id: string, method: string>
  child 0, source_file: string
  child 1, source_data_id: string
  child 2, method: string
to
{'schema_version': Value('string'), 'id': Value('string'), 'benchmark': Value('string'), 'ability': Value('string'), 'question': Value('string'), 'gt': Json(decode=True), 'input_token_num': Value('int64'), 'input_token_num_original': Value('int64'), 'input_token_num_mode': Value('string'), 'input': List({'role': Value('string'), 'content': Value('string')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              schema_version: string
              id: string
              benchmark: string
              ability: string
              question: string
              gt: string
              input_token_num: int64
              input_token_num_original: int64
              input_token_num_mode: string
              input: list<item: struct<role: string, content: string>>
                child 0, item: struct<role: string, content: string>
                    child 0, role: string
                    child 1, content: string
              longbenchv2_reconstruction: struct<source_file: string, source_data_id: string, method: string>
                child 0, source_file: string
                child 1, source_data_id: string
                child 2, method: string
              to
              {'schema_version': Value('string'), 'id': Value('string'), 'benchmark': Value('string'), 'ability': Value('string'), 'question': Value('string'), 'gt': Json(decode=True), 'input_token_num': Value('int64'), 'input_token_num_original': Value('int64'), 'input_token_num_mode': Value('string'), 'input': List({'role': Value('string'), 'content': Value('string')})}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

schema_version string	id string	benchmark string	ability string	question string	gt unknown	input_token_num int64	input_token_num_original int64	input_token_num_mode string	input list
v3_eval_v1	v3_mrcr_02b23e3d98b39461caa1cd52	MRCR	precise_retrieval	Prepend yoXLHgKpRG to the 6th (1 indexed) short scene in a play about elephants. Do not include any other text in your response.	{ "answer": "yoXLHgKpRGTitle: The Elephant in the Room\n\nScene 2: Elegance of Elephants\n\n*Setting: A peaceful savannah at dusk. The soft glow of the setting sun casts long shadows across the landscape. In the background, a shallow waterhole gleams. Center stage, beneath a large acacia tree, stand three ele...	7,120	7,237	qwen2.5-7b-instruct-chat-template	[ { "role": "system", "content": "You are a long-context reasoning assistant. The input contains [question], [Doc], and [BLOCK_ID: ...]. Output in this format:\n<evidence>{\"claim\": [{\"block_id\": \"...\", \"span\": \"start … end\"}]}</evidence>\n<summary>...</summary>\n<answer>...</answer>\n<evidence> and ...
v3_eval_v1	v3_mrcr_3ce096b92b3849e779499420	MRCR	precise_retrieval	"Prepend I2UtQoOC2J to the 1st (1 indexed) formal letter about stages. Do not include any other text(...TRUNCATED)	{"answer":"I2UtQoOC2J[Your Name] \n[Your Address] \n[City, State, ZIP Code] \n[Email Address] \n(...TRUNCATED)	103,034	106,113	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)
v3_eval_v1	v3_mrcr_84f63a2936d7ef9b74f2b41c	MRCR	precise_retrieval	"Prepend on8M92yVXV to the 3rd (1 indexed) riddle about systems. Do not include any other text in yo(...TRUNCATED)	{"answer":"on8M92yVXVI am a network, vast and wide, \nConnecting parts on every side. \nI’m the (...TRUNCATED)	107,821	110,932	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)
v3_eval_v1	v3_mrcr_0d5959dd839845523ac78155	MRCR	precise_retrieval	"Prepend Oa2JLAwFxQ to the 1st (1 indexed) song about afternoons. Do not include any other text in y(...TRUNCATED)	{"answer":"Oa2JLAwFxQVerse 1 \nWhen the morning light fades and the sun rises high, \nI find m(...TRUNCATED)	16,183	16,632	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)
v3_eval_v1	v3_mrcr_0fbf491898612c43806e2ac1	MRCR	precise_retrieval	"Prepend eJT7hYwPjf to the 6th (1 indexed) formal letter about materials. Do not include any other t(...TRUNCATED)	{"answer":"eJT7hYwPjf[Your Name] \n[Your Address] \n[City, State, ZIP Code] \n[Email Address] \n(...TRUNCATED)	29,191	29,929	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)
v3_eval_v1	v3_mrcr_02087ff8483e2598be0c0ec7	MRCR	precise_retrieval	"Prepend AyPc4if82I to the 4th (1 indexed) short news article about processes. Do not include any ot(...TRUNCATED)	{"answer":"AyPc4if82I**Title: Revolutionizing Efficiency: How Modern Processes Are Transforming Indu(...TRUNCATED)	30,766	31,590	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)
v3_eval_v1	v3_mrcr_03485534434f7509f9deae99	MRCR	precise_retrieval	"Prepend xx7ek7Tn7F to the 8th (1 indexed) short essay about studies. Do not include any other text (...TRUNCATED)	{"answer":"xx7ek7Tn7FTitle: The Enduring Value of Studies in Personal and Societal Growth\n\nIn an e(...TRUNCATED)	38,200	39,239	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)
v3_eval_v1	v3_mrcr_1b1db7aae7c0263b1b7767ce	MRCR	precise_retrieval	"Prepend 0cxoJ035Kz to the 1st (1 indexed) diary entry about tape. Do not include any other text in (...TRUNCATED)	{"answer":"0cxoJ035KzOctober 18, 2023\n\nDear Diary,\n\nToday I found myself thinking quite a bit ab(...TRUNCATED)	48,787	50,157	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)
v3_eval_v1	v3_mrcr_1a6aaff4d2c8b5bcdbb03f13	MRCR	precise_retrieval	"Prepend H4xYOWSVIw to the 1st (1 indexed) diary entry about walls. Do not include any other text in(...TRUNCATED)	{"answer":"H4xYOWSVIwDate: October 15, 2023\n\nDear Diary,\n\nToday, I found myself pondering th(...TRUNCATED)	56,047	57,554	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)
v3_eval_v1	v3_mrcr_0c1c4ff082acf03d287549e4	MRCR	precise_retrieval	"Prepend 31Tl5kD8oI to the 8th (1 indexed) email about crayons. Do not include any other text in you(...TRUNCATED)	{"answer":"31Tl5kD8oISubject: Exploring the Colorful World of Crayons!\n\nHi [Recipient's Name],\n\n(...TRUNCATED)	64,355	66,165	qwen2.5-7b-instruct-chat-template	[{"role":"system","content":"You are a long-context reasoning assistant. The input contains [questio(...TRUNCATED)

End of preview.

Highlight-Then-Summarize

Highlight-Then-Summarize (H2S) is a compress-then-reason approach for long-context understanding. It makes evidence localization and information integration explicit before final-answer generation:

long document + question -> evidence -> question-conditioned summary -> answer

[Code] [Data format]

H2S highlights source-addressable evidence from a block-structured document, integrates the selected information into a question-conditioned summary, and then generates the answer in one autoregressive trajectory. The original document remains available throughout generation: H2S organizes the context rather than pruning it.

What's Included

This repository releases two training sets and one evaluation benchmark:

Config	Split	File	Examples	Description
`sft`	`train`	`data/H2S-Dataset-SFT.jsonl`	4,228	Supervised Evidence--Summary--Answer trajectories
`rl`	`train`	`data/H2S-Dataset-RL.jsonl`	2,419	RL prompts with evidence, summary, and reward references
`bench`	`test`	`data/H2S-Bench.jsonl`	2,575	Seven-task long-context evaluation suite

Together, H2S-SFT and H2S-RL form H2S-Dataset, containing 6,647 examples from 11 benchmark families with an average context length of 43.9K tokens.

H2S-Bench

H2S-Bench evaluates seven complementary long-context capabilities. All models in the paper use the same block-structured inputs, a maximum input length of 128K tokens, and a maximum output length of 4K tokens. Only the final answer is used for task performance; generated evidence and summaries do not contribute to the benchmark score.

Benchmark	Capability	Examples	Evaluation
DocFinQA	Numerical reasoning	163	Numeric matching with tolerance
Frames	Multi-document reasoning	296	Exact match after answer normalization
LongCite	Citation-grounded answering	500	Harmonic mean of answer-content and citation F1
MRCR	Precise retrieval	500	Normalized character match at a 0.95 threshold
AA-LCR	Distributed-information aggregation	200	Native QA/set matching by answer contract
LongBenchV2	General long-context reasoning	416	Multiple-choice accuracy
HELMET-Summ	Long-document summarization	500	ROUGE-L
Total		2,575	Macro-average over the seven tasks

The packaged gt field contains benchmark-native final-answer references. H2S-Bench does not provide reference H2S evidence or reference summaries.

Dataset Construction

H2S-Dataset augments long-context question-answer instances with source-grounded evidence and a question-conditioned summary. Construction consists of semantic blocking, candidate retrieval, query-guided evidence extraction, summary synthesis, and quality verification. Retrieval is used only during offline data construction; H2S does not require an external retriever at inference time.

H2S-RL applies a deterministic process reward to answer quality, evidence grounding and span coverage, summary quality, and output structure, without an online LLM judge.

Usage

from datasets import load_dataset

sft = load_dataset("H2S-Research/Highlight-Then-Summarize", "sft", split="train")
rl = load_dataset("H2S-Research/Highlight-Then-Summarize", "rl", split="train")
bench = load_dataset("H2S-Research/Highlight-Then-Summarize", "bench", split="test")

For exact field definitions and JSONL examples, see data/README.md. Evaluation code is provided in the GitHub repository.

Main Results

All models use a maximum input length of 128K and a maximum output length of 4K. Scores are on a 0--100 scale, and Average is the unweighted macro-average over the seven tasks.

Category	Model	DocFinQA	Frames	LongCite	MRCR	AA-LCR	LongBenchV2	HELMET-Summ	Average
Closed-source	Claude Opus 5	56.11	27.89	12.17	100.00	26.45	64.18	15.13	43.13
Closed-source	GPT-5.5	49.57	36.05	12.62	92.38	52.34	46.15	9.45	42.65
Closed-source	DeepSeek-V4-Pro	34.83	37.41	6.78	72.14	47.77	48.56	12.48	37.14
Closed-source	Kimi-K2.5	51.14	37.41	10.83	87.58	39.95	50.00	14.41	41.62
Closed-source	GLM-5.2	49.57	36.05	12.62	92.38	40.85	48.80	16.52	42.40
Open-source	DeepSeek-R1-Distill-Qwen-7B	0.06	8.50	1.31	0.40	0.06	1.92	5.29	2.51
Open-source	DeepSeek-R1-Distill-Qwen-14B	11.69	23.13	5.25	38.28	8.75	23.32	11.81	17.46
Open-source	Qwen3-4B-Thinking-2507	14.61	18.71	3.68	25.45	9.88	18.99	10.80	14.59
Open-source	Qwen3.8-27B	40.52	32.31	4.94	46.89	15.69	15.14	1.51	22.43
Open-source	Qwen3.5-35B-A3B	24.72	18.03	2.13	8.62	6.11	6.97	0.06	9.52
Open-source	QwenLong-L1-32B	35.03	34.35	8.16	40.28	26.20	37.98	11.13	27.59
Open-source	QwenLong-L1.5-30B-A3B	9.34	25.17	1.03	0.00	5.37	30.52	3.92	10.76
Open-source	Qwen2.5-7B-Instruct-1M	5.00	13.95	5.37	4.01	7.31	24.04	7.16	9.55
Open-source	Qwen2.5-14B-Instruct-1M	15.70	23.47	8.32	25.05	13.56	35.82	12.03	19.14
Ours	H2S-7B-SFT	20.45	31.29	8.65	58.12	13.96	27.88	5.42	23.68
Ours	H2S-7B	26.33	32.97	15.76	59.72	21.47	32.11	11.58	28.56
Ours	H2S-14B-SFT	31.49	37.76	10.50	65.13	22.40	28.37	8.45	29.16
Ours	H2S-14B	30.83	37.72	13.88	66.13	26.86	35.42	17.38	32.60

H2S-RL improves over the SFT checkpoints by 4.88 points for 7B and 3.44 points for 14B. H2S-14B obtains the strongest average among the evaluated open-source models.

License and Citation

The records are derived from multiple upstream benchmarks. Users must follow the licenses and terms of the corresponding source datasets. The source-code license does not automatically apply to redistributed benchmark content.

Paper citation metadata will be added when the paper is publicly released.

Downloads last month: 49