metadata
pretty_name: My Custom Benchmark
dataset_info:
- config_name: chunked
features:
- name: document_id
dtype: string
- name: document_text
dtype: string
- name: document_filename
dtype: string
- name: document_metadata
struct:
- name: file_size
dtype: int64
- name: document_summary
dtype: string
- name: summarization_model
dtype: string
- name: chunks
list:
- name: chunk_id
dtype: string
- name: chunk_text
dtype: string
- name: multihop_chunks
list:
- name: chunk_ids
list: string
- name: chunks_text
list: string
splits:
- name: train
num_bytes: 635911
num_examples: 1
download_size: 243267
dataset_size: 635911
- config_name: ingested
features:
- name: document_id
dtype: string
- name: document_text
dtype: string
- name: document_filename
dtype: string
- name: document_metadata
struct:
- name: file_size
dtype: int64
splits:
- name: train
num_bytes: 133617
num_examples: 1
download_size: 78968
dataset_size: 133617
- config_name: multi_hop_questions
features:
- name: document_id
dtype: string
- name: additional_instructions
dtype: string
- name: question
dtype: string
- name: self_answer
dtype: string
- name: estimated_difficulty
dtype: int64
- name: self_assessed_question_type
dtype: string
- name: generating_model
dtype: string
- name: thought_process
dtype: string
- name: raw_response
dtype: string
- name: citations
list: string
- name: original_question
dtype: 'null'
- name: question_rewriting_model
dtype: 'null'
- name: question_rewriting_rationale
dtype: 'null'
- name: raw_question_rewriting_response
dtype: 'null'
- name: source_chunk_ids
list: string
splits:
- name: train
num_bytes: 216630
num_examples: 18
download_size: 38811
dataset_size: 216630
- config_name: prepared_lighteval
features:
- name: question
dtype: string
- name: additional_instructions
dtype: string
- name: ground_truth_answer
dtype: string
- name: gold
list: string
- name: choices
list: 'null'
- name: question_category
dtype: string
- name: kind
dtype: string
- name: estimated_difficulty
dtype: int64
- name: citations
list: string
- name: document_id
dtype: string
- name: chunk_ids
list: string
- name: question_generating_model
dtype: string
- name: chunks
list: string
- name: document
dtype: string
- name: document_summary
dtype: string
splits:
- name: train
num_bytes: 10018308
num_examples: 54
download_size: 204454
dataset_size: 10018308
- config_name: single_shot_questions
features:
- name: document_id
dtype: string
- name: additional_instructions
dtype: string
- name: question
dtype: string
- name: self_answer
dtype: string
- name: estimated_difficulty
dtype: int64
- name: self_assessed_question_type
dtype: string
- name: generating_model
dtype: string
- name: thought_process
dtype: string
- name: raw_response
dtype: string
- name: citations
list: string
- name: original_question
dtype: 'null'
- name: question_rewriting_model
dtype: 'null'
- name: question_rewriting_rationale
dtype: 'null'
- name: raw_question_rewriting_response
dtype: 'null'
- name: chunk_id
dtype: string
splits:
- name: train
num_bytes: 417825
num_examples: 36
download_size: 49046
dataset_size: 417825
- config_name: summarized
features:
- name: document_id
dtype: string
- name: document_text
dtype: string
- name: document_filename
dtype: string
- name: document_metadata
struct:
- name: file_size
dtype: int64
- name: document_summary
dtype: string
- name: summarization_model
dtype: string
splits:
- name: train
num_bytes: 133650
num_examples: 1
download_size: 79944
dataset_size: 133650
configs:
- config_name: chunked
data_files:
- split: train
path: chunked/train-*
- config_name: ingested
data_files:
- split: train
path: ingested/train-*
- config_name: multi_hop_questions
data_files:
- split: train
path: multi_hop_questions/train-*
- config_name: prepared_lighteval
data_files:
- split: train
path: prepared_lighteval/train-*
- config_name: single_shot_questions
data_files:
- split: train
path: single_shot_questions/train-*
- config_name: summarized
data_files:
- split: train
path: summarized/train-*
My Custom Benchmark
This dataset was generated using YourBench (v0.6.0), an open-source framework for generating domain-specific benchmarks from document collections.
Pipeline Steps
- ingestion: Read raw source documents, convert them to normalized markdown and save for downstream steps
- summarization: Perform hierarchical summarization: chunk-level LLM summaries followed by combine-stage reduction
- chunking: Split texts into token-based single-hop and multi-hop chunks
- single_shot_question_generation: Generate standalone question-answer pairs per chunk using LLM
- multi_hop_question_generation: Generate multi-hop QA pairs requiring reasoning across multiple chunks
- lighteval: Merge QA pairs and chunk metadata into a lighteval compatible dataset for quick model-based scoring
Reproducibility
To reproduce this dataset, use YourBench v0.6.0 with the following configuration:
hf_configuration:
hf_dataset_name: my_custom_benchmark
hf_organization: Tooru-Tooru
hf_token: $HF_TOKEN
local_dataset_dir: data/saved_dataset
jsonl_export_dir: data/jsonl_export
pipeline_config:
ingestion:
source_documents_dir: data/raw
output_dir: data/processed
pdf_llm_prompt: yourbench/prompts/ingestion/pdf_llm_prompt.md
summarization:
summarization_user_prompt: yourbench/prompts/summarization/summarization_user_prompt.md
combine_summaries_user_prompt: yourbench/prompts/summarization/combine_summaries_user_prompt.md
chunking: {}
single_shot_question_generation:
additional_instructions: Generate questions to test a curious adult
single_shot_system_prompt: yourbench/prompts/question_generation/single_shot_system_prompt.md
single_shot_system_prompt_multi: yourbench/prompts/question_generation/single_shot_system_prompt_multi.md
single_shot_user_prompt: yourbench/prompts/question_generation/single_shot_user_prompt.md
multi_hop_question_generation:
additional_instructions: Generate questions to test a curious adult
multi_hop_system_prompt: yourbench/prompts/question_generation/multi_hop_system_prompt.md
multi_hop_system_prompt_multi: '<custom_prompt: # Multi-Hop Document Comprehension
Question Genera...>'
multi_hop_user_prompt: yourbench/prompts/question_generation/multi_hop_user_prompt.md
lighteval: {}
model_list:
- model_name: Qwen/Qwen2.5-VL-72B-Instruct
api_key: $API_KEY
max_concurrent_requests: 32
encoding_name: cl100k_base
provider: hf-inference
- model_name: Qwen/Qwen2.5-72B-Instruct
api_key: $API_KEY
max_concurrent_requests: 32
encoding_name: cl100k_base
provider: novita
model_roles:
ingestion:
- Qwen/Qwen2.5-VL-72B-Instruct
summarization:
- Qwen/Qwen2.5-72B-Instruct
chunking:
- intfloat/multilingual-e5-large-instruct
single_shot_question_generation:
- Qwen/Qwen2.5-72B-Instruct
multi_hop_question_generation:
- Qwen/Qwen2.5-72B-Instruct
question_generation:
- Qwen/Qwen2.5-VL-72B-Instruct
cross_document_question_generation:
- Qwen/Qwen2.5-VL-72B-Instruct
question_rewriting:
- Qwen/Qwen2.5-VL-72B-Instruct
prepare_lighteval:
- Qwen/Qwen2.5-VL-72B-Instruct
lighteval:
- Qwen/Qwen2.5-VL-72B-Instruct
citation_score_filtering:
- Qwen/Qwen2.5-VL-72B-Instruct
(This dataset card was automatically generated by YourBench)