JesseLiu/agentorch-training

AgentOrch training split

Training data for global workflow orchestration from the paper AgentOrch: How Close Can Global Orchestration Come to Step-by-Step Interaction?. Each example pairs a task specification (the same Global View prompt format used by the AgentOrch benchmark, with a 64-card tool surface) with the complete serialized reference workflow of one source recipe. The 3,000 source recipes are disjoint from the 650 benchmark SOPs by recipe id; both are drawn from the same enterprise automation environment, so the tool ecosystem is shared while the recipes are unseen.

Targets keep the full workflow structure: both outcomes of conditionals, loop bodies, try/catch regions, tool calls, and references from later steps to runtime inputs and earlier results. Runtime variants are not used. Across the set the reference workflows use 1,033 tool operations over 305 namespaces. Every target passed structural and interface validation (syntax, node references, tool availability, argument schemas) before inclusion.

Benchmark prompts: JesseLiu/agentorch-benchmark. Code, scorer and the exporter that regenerates these files byte-for-byte: https://github.com/jesseliu0913/AgentOrch

Configurations

config	rows	contents
sft_messages_v1	3,000	canonical SFT export: messages (system, user, assistant), metadata (recipe id, hashes, split), schema_version
sft_v1_80k	2,789 train / 60 eval	the SFT set actually trained on: rows that fit an 81,920-token budget under the Qwen3.8-27B tokenizer (151 over-budget recipes dropped, listed in sft_v1_80k/prep_report.json)
rl_verl_v1	1,000 train / 47 eval	GRPO prompts for the verl route (Qwen models): prompt, reward_model holding the serialized reference workflow, extra_info
rl_trl_v1	500 train / 24 eval	GRPO prompts for the TRL route (Gemma 12B / 26B): prompt messages and golden reference workflow
rl_trl_v1_31b	131 train / 1 eval	shorter-prompt subset used for the Gemma 31B RL run

The RL prompt sets were selected by prompt length to fit the configured generation limits. source_manifest_v1.json and source_manifest_v2.json record the export settings and checksums.

Training recipes

Supervised fine-tuning applies LoRA (r=16, alpha=32) with the loss on the workflow response only. Reinforcement learning uses GRPO with eight samples per prompt and a reward that compares the sampled workflow with the reference (validity 0.15, tool-call sequence 0.25, arguments 0.35, structure 0.15, exact match 0.10) without executing anything. Trained checkpoints are published as JesseLiu/agentorch-sft-* and JesseLiu/agentorch-rl-* model repositories.

Source recipes (source/)

The 3,000 training recipes in raw form, packed as archives that unpack to their original paths in the code repository. Fetch and verify them with python scripts/download_data.py --only training (add --prepared for the SFT / RL files above).

archive	unpacks to	files
source/training_input_v1.tar.gz	training/input/recipes/ (v1 prompts, used by every reported run)	9,000
source/training_groundtruth.tar.gz	training/groundtruth/recipes/ (reference workflows)	6,000

Citation

@article{agentorch2026,
  title   = {AgentOrch: How Close Can Global Orchestration Come to Step-by-Step Interaction?},
  author  = {Liu, Zijie and Huang, Qi Yang and Duan, Jinhao and Lee, Yueh Che and Joseph, Thomas A. and Turner, Zayne and Seligman, Adam and Chen, Tianlong and Xu, Zhaozhuo},
  year    = {2026}
}