case_id string | benchmark_family string | question string | category string | answerability string | passages list | valid_evidence_sets list | reference_answer string | split string |
|---|---|---|---|---|---|---|---|---|
web-01 | web_5page | How does on-policy distillation differ from off-policy knowledge distillation and reinforcement learning? | direct | answerable | [
{
"id": 1,
"text": "A curated collection of papers, technical reports, frameworks, and tools for on-policy distillation (OPD) of large language models.\n\n> **On-policy distillation** trains a student on samples from its own evolving policy, while a teacher (external, privileged, or self-conditioned) provid... | [
[
1
]
] | On-policy distillation trains the student on its own generations, closing the train-inference distribution gap, and uses dense token-level teacher guidance rather than sparse rewards [1]. | real_web |
web-02 | web_5page | What objective does TailRL optimize, and what kinds of rollouts does its gradient emphasize? | paraphrase | answerable | [
{
"id": 1,
"text": "A curated collection of papers, technical reports, frameworks, and tools for on-policy distillation (OPD) of large language models.\n\n> **On-policy distillation** trains a student on samples from its own evolving policy, while a teacher (external, privileged, or self-conditioned) provid... | [
[
2
]
] | TailRL maximizes the log-probability that reward exceeds a randomly selected threshold, and its gradient emphasizes rare, high-reward rollouts [2]. | real_web |
web-03 | web_5page | What different roles do Tinker and OpenEnv play in an agentic reinforcement-learning workflow? | multi_source | answerable | [
{
"id": 1,
"text": "A curated collection of papers, technical reports, frameworks, and tools for on-policy distillation (OPD) of large language models.\n\n> **On-policy distillation** trains a student on samples from its own evolving policy, while a teacher (external, privileged, or self-conditioned) provid... | [
[
3,
4
]
] | Tinker supplies the model fine-tuning SDK and handles distributed-training complexity behind API requests [3]. OpenEnv supplies isolated execution environments with a standard Gymnasium-style `step()`, `reset()`, and `state()` interface [4]. | real_web |
web-04 | web_5page | Which exact public-cloud provider does the ML SYS tutorial recommend for reinforcement-learning infrastructure? | missing_information | unanswerable | [
{
"id": 1,
"text": "A curated collection of papers, technical reports, frameworks, and tools for on-policy distillation (OPD) of large language models.\n\n> **On-policy distillation** trains a student on samples from its own evolving policy, while a teacher (external, privileged, or self-conditioned) provid... | [] | The supplied passages do not provide enough information. | real_web |
web-05 | web_5page | How does Aleph avoid putting large files directly into an LLM prompt? | direct | answerable | [
{
"id": 1,
"text": "**Your RAM is the new context window.**\n\nAleph is an MCP server for recursive LLM workflows. Instead of forcing large files into prompt context, Aleph keeps data in a Python process and gives the model tools to inspect and reason over it.\n\n- Load big files and codebases into external... | [
[
1
]
] | Aleph keeps large-file data in a Python process outside the prompt and gives the model tools to inspect and reason over that external memory [1]. | real_web |
web-06 | web_5page | What problem was Craft Agents built to solve, and what principle does agent-native architecture require for user-interface capabilities? | multi_source | answerable | [
{
"id": 1,
"text": "**Your RAM is the new context window.**\n\nAleph is an MCP server for recursive LLM workflows. Instead of forcing large files into prompt context, Aleph keeps data in a Python process and gives the model tools to inspect and reason over it.\n\n- Load big files and codebases into external... | [
[
2,
4
]
] | Craft Agents was built so the Craft team could work effectively with agents [2]. Agent-native architecture requires that every outcome available through the UI also be achievable by the agent through tools [4]. | real_web |
web-07 | web_5page | Which prompt-template frontmatter fields does the Pi extension add, and what happens when the specialized mode finishes? | direct | answerable | [
{
"id": 1,
"text": "**Your RAM is the new context window.**\n\nAleph is an MCP server for recursive LLM workflows. Instead of forcing large files into prompt context, Aleph keeps data in a Python process and gives the model tools to inspect and reason over it.\n\n- Load big files and codebases into external... | [
[
3
]
] | The extension adds `model`, `skill`, and `thinking` frontmatter, then automatically restores the prior model when the specialized mode ends [3]. | real_web |
web-08 | web_5page | What four-step interaction loop does the agent-browser guide recommend after opening a page? | distractor | answerable | [
{
"id": 1,
"text": "agent-browser\n\nBrowser automation using Vercel's agent-browser CLI. Use when you need to interact with web pages, fill forms, take screenshots, or scrape data. Alternative to Playwright MCP - uses Bash commands with ref-based element selection. Triggers on \"browse website\", \"fill fo... | [
[
1
]
] | The guide recommends navigating to the URL, taking a snapshot to obtain element references, interacting through those references, and taking another snapshot after navigation or DOM changes [1]. | real_web |
web-09 | web_5page | What are Gemma 3n and FunctionGemma each designed for? | direct | answerable | [
{
"id": 1,
"text": "## Welcome to the Gemma Cookbook\n\nThis is a collection of guides and examples for Google Gemma.\n\n> **Disclaimer:** Gemma is a family of developer-focused models built by Google Deepmind. This cookbook is a collection of guides and examples for Google Gemma. Please keep in mind that G... | [
[
1
]
] | Gemma 3n is designed for efficient multimodal operation on low-resource devices, while FunctionGemma is a Gemma 3 270M instruction-tuned checkpoint specialized for function calling [1]. | real_web |
web-10 | web_5page | Which major components does GLiNKER combine into its entity-linking pipeline? | paraphrase | answerable | [
{
"id": 1,
"text": "## Welcome to the Gemma Cookbook\n\nThis is a collection of guides and examples for Google Gemma.\n\n> **Disclaimer:** Gemma is a family of developer-focused models built by Google Deepmind. This cookbook is a collection of guides and examples for Google Gemma. Please keep in mind that G... | [
[
2
]
] | GLiNKER combines named-entity recognition, multi-layer database search, and neural entity disambiguation in one modular pipeline [2]. | real_web |
web-11 | web_5page | What two forms of user representation does TokenMinds generate, and why retain both? | direct | answerable | [
{
"id": 1,
"text": "## Welcome to the Gemma Cookbook\n\nThis is a collection of guides and examples for Google Gemma.\n\n> **Disclaimer:** Gemma is a family of developer-focused models built by Google Deepmind. This cookbook is a collection of guides and examples for Google Gemma. Please keep in mind that G... | [
[
3
]
] | TokenMinds generates discrete SID-based user tokens and dense user embeddings; the tokens provide semantically grounded discrete representations while the embeddings preserve compatibility with downstream models [3]. | real_web |
web-12 | web_5page | What are information theory's two central coding problems, and which one concerns communication through noise? | direct | answerable | [
{
"id": 1,
"text": "## Welcome to the Gemma Cookbook\n\nThis is a collection of guides and examples for Google Gemma.\n\n> **Disclaimer:** Gemma is a family of developer-focused models built by Google Deepmind. This cookbook is a collection of guides and examples for Google Gemma. Please keep in mind that G... | [
[
4
]
] | The two problems are source coding for data compression and channel coding for data transmission; channel coding is the one concerned with recovering data sent through a noisy channel [4]. | real_web |
rag-01 | rag_10chunk | How can taxonomy child nodes, parents, and grandparents provide semantic behavior in a BM25 index, and which bag-of-words limitations remain for synonyms? | cross_document | answerable | [
{
"id": 1,
"text": "In the past, the management of a taxonomy would be massive. It might require complex rules, mapping of query phrases to taxonomy nodes, and a team of labelers and experts to keep everything straight. This has long been an appeal of embeddings.\n\nBut LLMs sneakily supercharge these old s... | [
[
1,
2
]
] | BM25 provides fast lexical matching but does not understand synonyms, word order, or semantic intent [2]. An LLM can classify products and queries into a taxonomy, whose parent paths are indexed as tokens; rare, specific child nodes then outrank common ancestors [1]. | rag_10chunk |
rag-02 | rag_10chunk | What exact overall recall percentage did the 10,000-query MIMICS distribution experiment report? | partial_or_miss | unanswerable | [
{
"id": 1,
"text": "_This is Part 2 of a three-part series on building and evaluating large-scale retrieval._\n\n_Part 1_ _covered the operational reality of working at this scale. This post covers what happened when we started running queries. Part 3 examines what happens when semantic and lexical retrieva... | [] | The supplied passages do not provide enough information. | rag_10chunk |
Tiny Evidence Analyst benchmark
This dataset tests short answers with passage-level citations. It has a synthetic evaluator-attack layer plus web_5page and rag_10chunk real-data families.
The synthetic layer contains 12 cases and 24 hand-labeled candidate outputs. web_5page contains 12 cases and 60 passages. rag_10chunk contains 12 cases and 120 chunks produced by a frozen BM25 retrieval over six Karakeep-backed corpora. This is a small benchmark, not a statistically powered leaderboard.
Task contract
Given one question and 1 to 15 numbered passages, produce no more than four cited sentences. Put each citation after the claim that it supports. Use only passage IDs from the input.
If a required fact is missing, output exactly The supplied passages do not provide enough information.
Dataset Viewer
The default configuration contains 24 rows and nine compact columns. benchmark_family distinguishes the two 12-case families. Each passage exposes only its ID, text, source title, and source URL; hashes, retrieval scores, offsets, source metadata, and reviews stay in the supporting archive.
Download supporting/benchmark-artifacts.zip for the synthetic calibration cases, evaluator attacks, complete provenance, semantic reviews, hashes, schemas, source-fetch records, and verifier code. These supporting files do not appear as separate datasets or Viewer columns.
from datasets import load_dataset
benchmark = load_dataset(
"jayshah5696/tiny-evidence-analyst-benchmark",
split="test",
)
Files
data/benchmark.parquet: the only Viewer dataset, with 24 real-data cases.supporting/benchmark-artifacts.zip: calibration cases, evaluator attacks, canonical records, provenance, schemas, hashes, and verifier code.DATA_LICENSE.md: source attribution and per-source licenses.scripts/build_real_web_benchmark.py: the source-refresh tool.scripts/build_rag_10chunk_benchmark.py: the deterministic chunking and retrieval replay tool.src/: the deterministic verifier and calibration audit.
Real-web construction
Each real-web case has five passages from five distinct source documents and source groups. The passages come from complete Karakeep reader-view fetches, including paginated fetches. The build script removes image-only lines and Markdown link destinations, then joins complete text blocks.
The script measures passages with Qwen/Qwen3.5-0.8B at revision 2fc06364715b967f1860aea9cf38778875588b17. Passage lengths range from 417 to 549 tokens. The average is 516 tokens. The normalized full-page fetches range from 699 to 61,604 tokens and average 8,529 tokens across the 16 pages.
Every passage records its canonical URL, source license, Karakeep bookmark ID, crawl time, content version, full-page size, full-content hash, passage hash, and token count. Every required fact cites a verbatim span from a passage. The missing-information case has no evidence span and requires the exact abstention string.
RAG construction
Each RAG case stores exactly 10 chunks in actual retrieval order. The builder creates non-overlapping, paragraph-preserving chunks, measures them with the pinned Qwen tokenizer, and ranks them with bm25-regex-v0 (k1=1.2, b=0.75) using a deterministic source-and-chunk tie break plus a four-chunk-per-source cap. The 120 chunks range from 240 to 310 tokens and average 276.8; top-10 contexts range from 2,706 to 2,839 tokens and average 2,768.3. Every context spans three to five source pages.
The records distinguish whether the full corpus contains the answer from whether the retrieved context is complete. Ten contexts contain complete answerable evidence. One asks for a fact absent from the corpus and requires the exact abstention string. One retrieves two conflicting source claims and requires the answer to surface both with separate citations.
Calibration composition
The cases include four direct questions, two multi-source questions, two missing-information questions, one distractor case, one conflict case, and two paraphrase variants.
The invalid outputs cover a wrong claim with a valid citation, citation-only text, copied input, an invented citation ID, an uncited second claim, an irrelevant citation, blanket abstention, a guessed answer, mixed support, a hidden conflict, an answer longer than four sentences, and a conversational preamble.
Proof boundary
Code checks the output form, sentence limit, citation addresses, answerability behavior, and evidence-set coverage. Human labels determine whether each claim is supported, required facts are covered, and conflicts are handled correctly. Citation validity alone does not prove claim support.
The semantic reviews are separate from the candidate verdicts, so the evaluator does not derive semantic success from the expected label. The first reviews are author-reviewed and explicitly marked as not blinded. Obtain independent, blinded review before using later development or test splits for model claims.
Run the included verifier from the repository root:
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python3 -m cited_answer_lab.audit --benchmark-dir data
Licenses
The benchmark code, schemas, questions, labels, reference answers, and synthetic calibration records use Apache-2.0. Web text keeps its original rights and terms. Several RAG sources do not declare an open reuse license, so the dataset license is other; consult DATA_LICENSE.md and each source URL before redistributing excerpts.
- Downloads last month
- 3