datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
BlueprintTrace
BlueprintTrace
142 failed Lean proof blueprints over 141 miniF2F targets.
Benchmark for BlueprintRepair: Typed Local Edits for Failed Lean Proof Blueprints,
accepted at MathNLP 2026 (EMNLP).
Ruslan Khrulev · Lomonosov Moscow State University
Code and evaluator · Paper · Citation
Benchmark
from datasets import load_dataset
states = load_dataset("Karifannaa/BlueprintTrace", "states", split="test")
Each row includes the broken Lean module, target statement, source… See the full description on the dataset page: https://huggingface.co/datasets/Karifannaa/BlueprintTrace.BluePrint
📘 BluePrint
BluePrint is a large-scale dataset of social media conversation threads designed for evaluating and training LLM-based social media agents.
It provides realistic, thread-structured data clustered into representative user personas at various levels of granularity.
✅ Key Features
Thread-Based Structure: Each example is a list of messages representing a user thread.
Persona Clustering: Users are clustered into 2, 25, 100, and 1000 representative personas to… See the full description on the dataset page: https://huggingface.co/datasets/ComplexDataLab/BluePrint.mille-agent-blueprints
MILLE Agent Blueprints
This dataset contains 200 synthetic, contract-checked examples for evaluating
agents that plan machine-learning systems. Each JSONL row pairs a plain-language
ML system request with a MILLE-generated expected blueprint, a dataset profile
when sample CSV data is available, rubric criteria, and known failure modes.
What it demonstrates
Evaluation across six ML task families and 24 operational domains
Explicit contracts for task framing… See the full description on the dataset page: https://huggingface.co/datasets/AhmedMSLTI/mille-agent-blueprints.claude-code-agent-blueprints
Claude Code Agent Blueprints — 14-Agent Production Company
Real instruction files for 10 production Claude Code agents running a live AI software company (whoffagents.com). Each record is one agent's complete role definition: scope, KPIs, tools, escalation paths, and collaboration rules.
What's in here
A real company built on 14 Claude Code agents coordinating via PAX Protocol. This dataset releases the actual agent instruction files (AGENTS.md) for the VP-level agents:… See the full description on the dataset page: https://huggingface.co/datasets/WH0FF/claude-code-agent-blueprints.alpaca-gpt4-originalblueprint03