TaskTrove Agentic Tasks

## Overview TaskTrove is an open-source aggregation of over 750,000 unique agentic tasks drawn from more than 100 task sources, released by the OpenThoughts-Agent team. The collection is the task-side complement to AgentTrove (which contains traces of models executing these tasks via the Harbor framework). Distributed as parquet files, it covers popular reinforcement learning and supervised fine-tuning targets for code and agent training, including SWE-Smith, R2EGym, and SWE-Re-Bench. ## Schema Schema varies by source subset; common fields across task records typically include: - `task_id` — string — unique task identifier - `source` — string — originating dataset/benchmark name - `instruction` / `problem_statement` — string — the task prompt or issue description - `repo` — string — source repository (for SWE-style tasks) - `base_commit` — string — git commit hash for environment setup - `test_patch` — string — evaluation patch/tests (where applicable) - `gold_patch` — string — reference solution (where available) - `environment_setup` — string/json — setup instructions or container spec - `metadata` — json — source-specific extra fields - See the dataset page for per-source variations across 100+ sources ## Sources - HuggingFace: https://huggingface.co/datasets/open-thoughts/TaskTrove — Apache-2.0 - Aggregates 100+ upstream task datasets including SWE-Smith, R2EGym, SWE-Re-Bench, and other agentic RL/SFT corpora (see dataset card for full list and upstream attributions) ## Methodology The OpenThoughts-Agent team curated and aggregated tasks from over 100 existing agentic task datasets used in RL and SFT pipelines, normalizing them into parquet shards organized by source. Each source subset preserves the upstream task format while enabling unified loading. Tasks were selected from popular training and evaluation targets for software engineering agents and general agentic workflows; AgentTrove (separate dataset) contains rollouts executed against these tasks using the Harbor framework. ## Known gaps & limitations Because TaskTrove aggregates 100+ heterogeneous sources, schema and task quality vary significantly between subsets. Many upstream sources (SWE-Bench, SWE-Re-Bench, etc.) are widely used in public agent evaluations — buyers using this for training should expect substantial leakage risk against popular benchmarks. Task validity and reproducibility (environment setup, container availability) depend on each upstream source. The dataset card does not exhaustively document per-source filtering, dedup, or quality scoring; buyers should validate task subsets empirically before training on them. ## Intended use & out-of-scope - **Intended use:** RL and SFT training for software engineering and general agentic models; task pool for generating new agent traces via Harbor or similar frameworks; research on agent task distributions. - **Out-of-scope:** Not deduplicated against common eval suites (SWE-Bench Verified, etc.) — direct training risks benchmark contamination. Not a curated quality-graded set; raw aggregation only. _Federated dataset: 5 parquet shards, 52.0 MB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: TaskTrove — 750K+ Agentic Tasks for RL & SFT Training Open-source collection of 750,000+ unique agentic tasks aggregated from 100+ sources including SWE-Smith, R2EGym, and SWE-Re-Bench. Apache-2.0 licensed, parquet format, designed for agent training and evaluation.