Datasets:

Max00035
/

rag-systems-eval-benchmark

Error code:   StreamingRowsError
Exception:    ValueError
Message:      Expected object or value
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 281, in _generate_tables
                  examples = [ujson_loads(line) for line in batch.splitlines()]
                              ~~~~~~~~~~~^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
                  return pd.io.json.ujson_loads(*args, **kwargs)
                         ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
              ValueError: Expected object or value

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.

RAG Systems Evaluation Bench

Dataset Summary

RAG Systems Evaluation Bench ragbench-v1 is a small, curated diagnostic dataset for comparing lexical BM25, dense MiniLM, and hybrid Reciprocal Rank Fusion retrieval. It contains original synthetic English technical passages about RAG and ML systems, explicit chunk-level relevance labels, and reference answers.

Dataset repository: Max00035/rag-systems-eval-benchmark
Static results dashboard: Max00035/rag-systems-eval-bench
Source code: mukeshbhandarkar/rag-systems-eval-bench

It is intended for engineering evaluation, education, and reproducibility demonstrations. It is not an academically validated or industry-standard benchmark, and included scores are not state-of-the-art claims.

Motivation

The benchmark makes retrieval behavior inspectable across exact lexical matches, paraphrases, acronyms, opaque identifiers, distractors, and straightforward controls. Persisted reference results show how retrieval failures can be distinguished from generation failures without an opaque LLM judge.

Dataset Structure

corpus.jsonl: 12 original source documents
chunks.jsonl: 36 deterministic chunks used for retrieval
benchmark.jsonl: 24 questions and annotations
retrieval_results.jsonl: precomputed BM25, Dense MiniLM, and Hybrid RRF rankings
benchmark_summary.json: aggregate and per-category metrics plus observed resource measurements

Fields

Benchmark records contain:

question_id: stable question identifier
question: English question text
relevant_chunk_ids: explicitly curated relevant chunks
reference_answers: one or more acceptable short answers
answerable: whether the synthetic corpus supports an answer
category: one of lexical, semantic, acronym, distractor, paraphrase, control, or unanswerable
metadata: reserved lightweight metadata

Benchmark Categories

Categories are diagnostic slices, not statistically powered sub-benchmarks. Each contains only two to four questions. They expose exact identifiers, semantic rewording, ambiguous vocabulary, acronyms, and basic controls.

Corpus Construction

Every corpus passage and benchmark question was created specifically for this benchmark as original synthetic, generic technical content. They were not scraped or copied from vendor documentation. Documents are deterministically divided into 45-word windows with 5-word overlap. Reference answers and relevance annotations were manually and explicitly selected only after stable chunk IDs were generated.

Retrieval Systems

BM25 with lowercase regex tokenization, k1=1.5, and b=0.75
sentence-transformers/all-MiniLM-L6-v2 with normalized embeddings and cosine-equivalent dot product
Hybrid Reciprocal Rank Fusion with rrf_k=60 and candidate_k=20

All reference runs use identical corpus, questions, labels, and cutoffs.

Evaluation Metrics

The summary includes Recall@1, Recall@3, Recall@5, and MRR at each cutoff. Reference generation uses BM25 with google/flan-t5-small; answer evaluation reports normalized exact match, token F1, optional semantic similarity, abstention, and failure-decomposition counts.

Exact match can underrate valid verbose answers. Token F1 measures surface overlap. Semantic similarity is not a faithfulness metric, and a semantically similar answer may be unsupported by retrieved context.

Reproduction

See the source project reproduction documentation. In short: sync with uv, rebuild and validate the corpus, run all three retrievers at top 5, then build the dashboard and dataset data from persisted artifacts. Generated results are presented in the static Space.

Limitations

This is a small synthetic diagnostic benchmark with manually curated labels and tiny category groups. It does not establish statistical significance, production performance, factual coverage of the broader ML field, or answer faithfulness. Labels may omit equivalent evidence. Runtime measurements describe one CPU-only development environment and are not universal performance claims.

Licensing

The original synthetic benchmark dataset is licensed under the Creative Commons Attribution 4.0 International License. See LICENSE in this dataset repository for the license reference and attribution notice. Source code and the static dashboard are separately licensed under MIT.

Versioning

Both the corpus and annotations use ragbench-v1. Any change to source text, chunk settings, questions, references, or relevance labels requires a reviewed version update.

Downloads last month: 16

Total file size:

105 kB