AgentTrove Agentic Interaction Traces

## Overview AgentTrove is the largest open-source collection of agentic interaction traces to date, released by the OpenThoughts-Agent team. It contains 1,696,847 rows drawn from 219 source datasets spanning code repair, shell scripting, mathematical problem-solving, competitive programming, and general computer-use tasks. At 1.7M rows, it is roughly 4× the size of the Nemotron Terminal Corpus (430K rows), the previous largest open-source agentic trace dataset. Distributed as Parquet, English-language, text modality. ## Schema The dataset is provided as Parquet with agentic interaction traces. Typical columns in agentic-trace corpora of this kind include: - `messages` / `trajectory` — list/JSON — multi-turn agent interaction with tool calls - `task` / `instruction` — string — original task prompt - `source_dataset` — string — which of the 219 source datasets the row came from - `category` — string — task category (code repair, shell, math, competitive programming, computer-use) - `tools` — list — tool/function schemas available to the agent - `outcome` / `success` — bool/string — whether the trajectory solved the task - `language` — string — primary language (en) - `length` / `num_turns` — int — trace length metadata Buyers should consult the HF dataset page for the authoritative schema as published. ## Sources - open-thoughts/AgentTrove on HuggingFace — https://huggingface.co/datasets/open-thoughts/AgentTrove — Apache 2.0 - Aggregated from 219 upstream source datasets (see dataset card for full attribution list) ## Methodology The OpenThoughts-Agent team aggregated agentic interaction traces from 219 existing open-source datasets covering code repair, shell scripting, math, competitive programming, and computer-use tasks. Traces were normalized into a unified schema and packaged as Parquet for scalable consumption. Some traces are originally machine-generated by agent frameworks (e.g., Terminus-2, Harbor) operating against benchmark tasks. ## Known gaps & limitations - English-only; no multilingual coverage. - Heterogeneous quality: aggregated from 219 sources with varying generation methods, model providers, and success criteria — trace quality is not uniform. - Likely leakage risk against common code/agent benchmarks (SWE-bench, HumanEval, MATH, etc.) since upstream sources may have been generated on those tasks. Not deduplicated against held-out eval suites. - Source dataset licenses vary; the Apache 2.0 license applies to the AgentTrove aggregation but downstream users should verify compatibility for individual source contributions if redistributing subsets. - Source does not exhaustively document per-row provenance gaps; buyers should validate empirically. ## Intended use & out-of-scope - **Intended:** SFT and RL fine-tuning of agentic / tool-using LLMs, behavior cloning, trajectory analysis, agent capability research, RAG over agent traces. - **Out-of-scope:** Training models for benchmark evaluation without first deduplicating against target eval sets (high leakage risk); production deployment without quality filtering across the 219 heterogeneous sources. _Federated dataset: 38 parquet shards, 18.21 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: AgentTrove: 1.7M Agentic Interaction Traces Largest open-source collection of agentic interaction traces (1.7M rows from 219 source datasets) covering code repair, shell scripting, math, competitive programming, and computer-use tasks. Apache 2.0.