Datasets:

jacobli
/

studybench

license: cc-by-4.0
language:
  - en
tags:
  - code
  - question-answering
  - llm-evaluation
  - rubric-grading
  - agents
task_categories:
  - question-answering
  - text-generation
size_categories:
  - n<1K
configs:
  - config_name: dspy
    data_files: dspy.jsonl
  - config_name: openclaw
    data_files: openclaw.jsonl

studybench

studybench is a small, high-effort benchmark of expert-level coding questions about real open-source codebases, each paired with a gold answer and a weighted, source-grounded grading rubric. The questions ask a model to produce working code that uses a specific library/framework correctly; the rubric decomposes a correct answer into discrete, checkable claims, each tied to exact lines of the upstream source.

This release publishes both the questions and the full rubrics (nothing is held back), so the evaluation is fully transparent and reproducible.

Configs

Pick a subset with the second argument of load_dataset:

from datasets import load_dataset

dspy     = load_dataset("jacobli/studybench", "dspy")        # 30 questions
openclaw = load_dataset("jacobli/studybench", "openclaw")    # 20 questions

config	questions	topics	codebase
`dspy`	30	6	DSPy
`openclaw`	20	4	OpenClaw

Schema

Each row has six fields:

field	type	description
`id`	string	stable opaque identifier
`topic`	string	coarse category (see below)
`question`	string	the task prompt — asks for a self-contained, runnable solution
`gold_answer`	string	a reference solution (code)
`rubric`	list	weighted claims that define a correct answer
`evidence`	list	source excerpts that ground the rubric

rubric — a list of claims; weights sum to 100 per question:

{
  "claim_id": "c1",
  "claim_type": "core",          // "core" = essential; "supporting" = secondary
  "weight": 52,                   // integer; the rubric's weights sum to 100
  "statement": "…what must be true of a correct answer…",
  "span_ids": ["s4", "s8"]        // evidence spans grounding this claim
}

evidence — the source excerpts the grader is shown; every span_ids value in rubric resolves to one of these span_ids:

{
  "span_id": "s4",
  "path": "dspy/teleprompt/gepa/gepa.py",   // path within the upstream repo
  "start_line": 330,
  "end_line": 365,
  "excerpt": "0330:     def __init__(\n0331:         self,\n…"   // line-number-prefixed source
}

Excerpts are byte-exact copies of the upstream source at the pinned commits below (each line is prefixed with its 1-indexed line number, e.g. 0330: ).

Topics

dspy: gepa_optimizer_usage, prompt_optimization_workflows, rag_and_retrieval_pipelines, react_agents_and_tools, signature_schema_and_pydantic_types, evaluation_metrics_and_custom_eval
openclaw: model_fallback_and_failover_logic, cross_session_channel_context_and_session_behavior_requests, memory_core_dreaming_and_promotion_pipeline, new_plugin_provider_and_channel_integration_requests

How the rubric is used for grading

A judge model is shown the question, the candidate answer, the gold_answer, the rubric, and the evidence spans. It scores each claim independently (does the answer satisfy the claim?), and the question score is the weight-weighted fraction of satisfied claims (0–100). claim_type lets you apply an optional conjunctive gate: require every core claim to be satisfied or the answer scores 0. The evidence excerpts are the only code context the judge needs — grading does not require checking out the repositories.

Source code & attribution

The evidence excerpts and path values reference these repositories at fixed commits:

codebase	repo	commit	license
DSPy	`stanfordnlp/dspy`	`9cdb0aac28b2a04b064e40697ccd301872cf6a43`	MIT
OpenClaw	`openclaw/openclaw`	`da228660306b55a9cce3b973946f3aacfc515848`	MIT

To inspect or extend the evidence, check out the corresponding repo at the pinned commit and open the listed path at the given line range.

Licensing

Questions, gold answers, and rubrics (the original contributions of this dataset) are released under CC-BY-4.0.
Embedded source excerpts are derived from DSPy and OpenClaw and remain under their respective MIT licenses; attribution is provided above.

Notes & limitations

This is a deliberately small, expert-curated set (50 questions total), not a large-scale benchmark.
Because both questions and rubrics are public, treat results as an open (non-held-out) evaluation; models may be trained on this content.
The benchmark is grounded in specific repository snapshots; answers reflect the APIs at the pinned commits.