source_index int64 0 49 | context_index int64 0 0 | question_index int64 0 0 | question large_stringlengths 129 2.96k | rubric large_stringlengths 0 1.08k | reference_answer large_stringlengths 5 744 | context large_stringlengths 101k 16.2M | context_tokens int64 24.4k 4.16M |
|---|---|---|---|---|---|---|---|
0 | 0 | 0 | "In the function that calculates the derivative of given functions, which of the following keyword a(...TRUNCATED) | "CORRECT: The response identifies choice (B) h, method, direction as the correct answer.\nWRONG: The(...TRUNCATED) | B: h, method, direction | "\"\"\"\nImplements the PSLQ algorithm for integer relation detection,\nand derivative algorithms fo(...TRUNCATED) | 783,531 |
1 | 0 | 0 | "I want to extend the task of Agentbench. My task is a mobile operation task, implemented using an A(...TRUNCATED) | C: Exit AVD in the release function and end testing Docker | "# AgentBench\n\n\n\n<p align=\"center\">\n <a href=\"https://llmbench.ai\"(...TRUNCATED) | 1,128,594 | |
2 | 0 | 0 | "I plan to use this framework to train the glm-4v-9b model. Which of the follwing operations will le(...TRUNCATED) | "D: After fine-tuning, I want to deploy the model service. I need to use swift infer --model_type gl(...TRUNCATED) | "# SWIFT (Scalable lightWeight Infrastructure for Fine-Tuning)\n\n<p align=\"center\">\n <br>\n (...TRUNCATED) | 1,165,649 | |
3 | 0 | 0 | "This is the troch.nn modeule. In this module, there exists an implementation of flexible attention (...TRUNCATED) | "• The response must identify (A) as the CORRECT choice, or explicitly state that the default BLOC(...TRUNCATED) | A: _DEFAULT_SPARSE_BLOCK_SIZE, _ModificationType.SCORE_MOD | "# mypy: allow-untyped-defs\n\"\"\"Functionality for Python <-> C++ frontend inter-op.\"\"\"\n\nfrom(...TRUNCATED) | 384,276 |
4 | 0 | 0 | "Which realistic factor in collaborative perception does this algorithm model mainly address?\n\nCho(...TRUNCATED) | "* The response must state that the algorithm addresses time asynchrony and posture (or pose) errors(...TRUNCATED) | "B: This model takes into account real-world problems, which are time asynchrony and posture errors,(...TRUNCATED) | "\"\"\"Specifies the current version number of v2xvit.\"\"\"\n\n__version__ = \"0.1.0\"\n\n\n\n\nimp(...TRUNCATED) | 76,649 |
5 | 0 | 0 | "The Instant3D paper introduced significant innovations in accelerating 3D object generation by redu(...TRUNCATED) | "• The correct answer is Choice (A), which states that OpenLRM's reliance on large datasets like O(...TRUNCATED) | "A: In the repo, OpenLRM’s reliance on large datasets like Objaverse and MVImgNet introduces chall(...TRUNCATED) | "# OpenLRM: Open-Source Large Reconstruction Models\n\n[ | 117,449 |
6 | 0 | 0 | "In the urls method of the Channel class, what does not determine the final URL list that is returne(...TRUNCATED) | "The response must identify Choice (A) as the correct answer.\n* The response is CORRECT if it state(...TRUNCATED) | "A: Handling of subdirs: If subdirs is not provided (i.e., None), the method assigns it the default (...TRUNCATED) | "from __future__ import annotations\n\nimport logging\nimport re\nimport sys\nfrom functools import (...TRUNCATED) | 441,327 |
7 | 0 | 0 | "In the FileManager class, which of the following wrongly describes the purpose of the write_with_te(...TRUNCATED) | "CORRECT responses will identify Choice (A) as the wrong description. The response should specifical(...TRUNCATED) | "A: Template Substitution Handling: The write_with_template method calls substitute_with_template to(...TRUNCATED) | "from __future__ import annotations\n\nimport argparse\nimport os\nimport re\nfrom collections impor(...TRUNCATED) | 281,159 |
8 | 0 | 0 | "What inputs are necessary to create an instance of HYPRE_SStructMatrix? And what procedures are nec(...TRUNCATED) | "A: An MPI communicator and a HYPRE_SStructGraph are necessary to create. HYPRE_SStructMatrixAssembl(...TRUNCATED) | "cmake_minimum_required(VERSION 3.13...3.16)\n\nif (${CMAKE_VERSION} VERSION_LESS 3.16)\n cmake_pol(...TRUNCATED) | 4,163,702 | |
9 | 0 | 0 | "In the May 20, 2023 commit of the trl repository, a new PPOTrainer class was introduced for Proxima(...TRUNCATED) | "The evaluator should determine if the response identifies the correct modification for preserving g(...TRUNCATED) | "D: Modify the training_step method in ppo_trainer.py to call accelerator.backward() at the end of e(...TRUNCATED) | "<div style=\"text-align: center\">\n<img src=\"https://huggingface.co/datasets/trl-internal-testing(...TRUNCATED) | 399,970 |
End of preview. Expand in Data Studio
LongBench Synthetic V4.1
Dataset statistics
v4.1 = lbs_v4 train+val (verbatim) + a test split absorbing every config from dac-research/extra_evals_v1 not already in v4. Five ZeroScrolls configs whose upstream corpora collide with lbs_v3.1 train/test (gov_report, qmsum, qasper, narrative_qa, musique) were dropped outright. All test rows have rubrics backfilled via the longbench_generate_rubric_for_imported.jinja template (same path as lbs_v3.1 / lbs_v4 Stage 1b). Token counts use Qwen/Qwen3-14B.
| Subset | Split | Unique ctx | Sample rows | <8K | 8-16K | 16-32K | 32-64K | 64-128K | >128K | Median tok | p90 tok | Max tok |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| longbench_v2_code_repo | test | 49 | 50 | 0 | 0 | 3 | 3 | 9 | 35 | 441,327 | 3,328,207 | 4,163,702 |
| longbench_v2_dialogue_history | test | 38 | 39 | 0 | 0 | 12 | 8 | 19 | 0 | 64,605 | 120,061 | 125,333 |
| longbench_v2_in_context_learning | test | 52 | 81 | 0 | 2 | 5 | 6 | 27 | 41 | 131,501 | 811,694 | 1,474,151 |
| longbench_v2_multidoc_qa | train | 110 | 880 | 0 | 96 | 128 | 264 | 136 | 256 | 59,953 | 243,558 | 960,383 |
| longbench_v2_multidoc_qa | validation | 123 | 984 | 0 | 104 | 152 | 264 | 152 | 312 | 61,425 | 270,817 | 960,383 |
| longbench_v2_singledoc_qa | train | 149 | 1,240 | 0 | 104 | 312 | 104 | 312 | 408 | 79,362 | 221,060 | 865,154 |
| longbench_v2_singledoc_qa | validation | 166 | 1,400 | 0 | 112 | 376 | 128 | 352 | 432 | 75,354 | 221,060 | 865,154 |
| longbench_v2_structured_data | test | 32 | 33 | 0 | 0 | 1 | 1 | 2 | 29 | 201,731 | 1,164,543 | 3,225,507 |
| loogle_longdep_qa | train | 126 | 994 | 0 | 158 | 336 | 495 | 5 | 0 | 33,230 | 45,133 | 83,038 |
| loogle_longdep_qa | validation | 140 | 1,101 | 0 | 158 | 391 | 547 | 5 | 0 | 33,230 | 44,972 | 83,038 |
| loogle_shortdep_cloze | test | 71 | 1,342 | 0 | 36 | 629 | 659 | 18 | 0 | 32,986 | 52,892 | 79,361 |
| loogle_shortdep_qa | train | 94 | 1,671 | 0 | 393 | 1,088 | 190 | 0 | 0 | 19,643 | 36,057 | 57,819 |
| loogle_shortdep_qa | validation | 105 | 1,951 | 0 | 468 | 1,293 | 190 | 0 | 0 | 19,643 | 32,268 | 57,819 |
| loogle_summarization | train | 459 | 3,672 | 0 | 1,200 | 2,224 | 208 | 32 | 8 | 17,879 | 29,769 | 146,446 |
| loogle_summarization | validation | 510 | 4,080 | 0 | 1,368 | 2,440 | 224 | 32 | 16 | 17,680 | 29,469 | 287,065 |
| ruler_cwe | test | 1,500 | 1,500 | 500 | 500 | 500 | 0 | 0 | 0 | 11,020 | 22,335 | 22,411 |
| ruler_fwe | test | 1,500 | 1,500 | 988 | 511 | 1 | 0 | 0 | 0 | 7,814 | 15,493 | 16,812 |
| ruler_niah_multikey_1 | test | 1,500 | 1,500 | 1,000 | 500 | 0 | 0 | 0 | 0 | 7,504 | 15,976 | 15,982 |
| ruler_niah_multikey_2 | test | 1,500 | 1,500 | 500 | 500 | 500 | 0 | 0 | 0 | 9,320 | 19,697 | 19,782 |
| ruler_niah_multikey_3 | test | 1,500 | 1,500 | 500 | 500 | 500 | 0 | 0 | 0 | 9,594 | 21,006 | 21,123 |
| ruler_niah_multiquery | test | 1,500 | 1,500 | 1,000 | 500 | 0 | 0 | 0 | 0 | 7,504 | 15,976 | 15,981 |
| ruler_niah_multivalue | test | 1,500 | 1,500 | 1,000 | 500 | 0 | 0 | 0 | 0 | 7,505 | 15,976 | 15,992 |
| ruler_niah_single_1 | test | 1,500 | 1,500 | 1,000 | 500 | 0 | 0 | 0 | 0 | 7,849 | 15,649 | 15,654 |
| ruler_niah_single_2 | test | 1,500 | 1,500 | 1,000 | 500 | 0 | 0 | 0 | 0 | 8,029 | 15,910 | 15,915 |
| ruler_niah_single_3 | test | 1,500 | 1,500 | 1,000 | 500 | 0 | 0 | 0 | 0 | 8,054 | 15,937 | 15,941 |
| ruler_qa_1 | test | 1,500 | 1,500 | 987 | 512 | 1 | 0 | 0 | 0 | 6,139 | 13,878 | 16,849 |
| ruler_qa_2 | test | 1,500 | 1,500 | 861 | 388 | 251 | 0 | 0 | 0 | 7,827 | 16,787 | 17,289 |
| ruler_vt | test | 1,500 | 1,500 | 1,000 | 500 | 0 | 0 | 0 | 0 | 7,878 | 16,279 | 16,285 |
| zeroscrolls_book_sum_sort | test | 520 | 520 | 292 | 196 | 32 | 0 | 0 | 0 | 7,533 | 14,367 | 23,831 |
| zeroscrolls_quality | test | 20 | 21 | 19 | 2 | 0 | 0 | 0 | 0 | 6,960 | 8,076 | 8,332 |
| zeroscrolls_space_digest | test | 520 | 520 | 514 | 6 | 0 | 0 | 0 | 0 | 6,751 | 7,531 | 8,355 |
| zeroscrolls_squality | test | 67 | 1,120 | 1,100 | 20 | 0 | 0 | 0 | 0 | 6,934 | 7,498 | 9,139 |
| zeroscrolls_summ_screen_fd | test | 357 | 357 | 179 | 172 | 6 | 0 | 0 | 0 | 8,151 | 13,134 | 23,341 |
Total: 22,210 unique contexts -> 33,099 sample rows.
Note on
validation: for every in-domain subset (those shippingtrain+validation) the HFvalidationsplit contains a pooled copy of that subset'strainrows in addition to the held-out validation rows, to grow the eval sample count for subsets with small validation holdouts (e.g.longbench_v2_multidoc_qa,loogle_longdep_qa). The per-split rows above reflect this pooling; the total is de-duplicated. Thetrainsplit is unchanged.
Provenance
- Train + validation: identical to
dac-research/longbench_synthetic_v4. - Test: from
dac-research/extra_evals_v1, with SHA-256(context) dedup against v4 train+val and lbs_v3.1 train+val+test. The collision-prone ZeroScrolls subsets above were excluded by config name (corpus-level overlap that hash dedup cannot detect). extra_evals_v1is deprecated as of v4.1; this dataset is the canonical successor.
- Downloads last month
- 347