CompassioninMachineLearning/olmo-part1-tool-data

Olmo part-one tool-use data (post-training)

The tool-use data used for part one (tool-use post-training) of the Olmo-3-7B anchors in the mid-training compassion experiments: olmo-urban-part1, olmo-urban-s1-part1, olmo-compassion-part1 and olmo-compassion-s1-part1. Part one turns a mid-trained base model into a tool-using agent before evaluation on TAC and HarvestBench.

This repository holds the parts we can redistribute. The full training mix is rebuilt by pilot/erosion/build_anchor.sh from the sources listed below.

Files

File	Rows	What it is
xlam_singlecall_5000.jsonl	5,000	Short single-turn requests, each answered with exactly one tool call. Derived from Salesforce/xlam-function-calling-60k (CC-BY-4.0) by pilot/build_xlam.py --n 5000 --max-calls 1 --seed 43: rows with more than one call, and rows mentioning bookings, tours, travel or animals, are dropped, so nothing resembles the evaluation tasks. SHA-256 a8aae6cf5c36dcb61b311fed62a8a5b61674e0aa6a3929ab7ddcbf4a7389b5cd, identical to the file every Olmo anchor trained on.
efficiency_500.jsonl	500	Synthetic booking dialogues that use a search / details / availability / purchase tool interface. No animal-related experiences. Shared by the Qwen and Olmo anchors.

Each row is {"messages": [...], "tools": [...], "source": ...} with structured tool_calls, so each model's own chat template renders the calls (Olmo: pythonic calls inside <function_calls>).

Full part-one mix (Olmo)

Source	Rows	Where
APIGen-MT multi-turn agent dialogues	4,824 after length filter	Salesforce/APIGen-MT-5k (gated, not redistributed here)
Tool-error recovery dialogues	3,750	CompassioninMachineLearning/agentic-tool-recovery-sft
Efficiency booking dialogues	500	this repository
xLAM single-call rows	5,000	this repository (Olmo only)
Own-domain replay	about 1,270 documents	the arm's mid-training corpus, formatted as an assistant reply to a neutral request and capped at 25% of supervised tokens

Training: LoRA r64 / alpha 128 with embeddings trained, 1-epoch cosine schedule, training seed 7 (8 for olmo-urban-s1-part1). Each anchor is the checkpoint at step 200, the earliest step at which all four Olmo anchors passed the capability gate (all 52 TAC samples, at least 90% completed with every scenario completed once, and at least 80% of HarvestBench answers parsed).

Why the xLAM rows (Olmo only)

The Olmo-3-7B base has no chat prior. With the multi-turn tool data alone it learned the call format inside those dialogues but rarely started a reply to a new request with a call, and wrote JSON calls the Olmo tool parser rejects. Short, varied single-call examples fixed that. Rows with two calls were tried first and taught Olmo to chain calls with guessed IDs before seeing results, so only single-call rows are used. The Qwen3-8B anchors do not use them.

License and citation

The xLAM-derived rows keep the CC-BY-4.0 license of the source, and this repository uses the same access acknowledgement. The source is released for research purposes. Please cite APIGen:

@article{liu2024apigen,
  title={APIGen: Automated Pipeline for Generating Verifiable and Diverse Function-Calling Datasets},
  author={Liu, Zuxin and Hoang, Thai and Zhang, Jianguo and Zhu, Ming and Lan, Tian and Kokane, Shirley and Tan, Juntao and Yao, Weiran and Liu, Zhiwei and Feng, Yihao and others},
  journal={arXiv preprint arXiv:2406.18518},
  year={2024}
}