Tholos trajectories: tool-use training data for small AI agents
0
7 commits
1 linked in READMEs
updated Oct 9, 2026
Executed English conversations for training small language models to call tools, follow approval rules and update a shared workspace. This public subset helped train Tholos-2B, which passes 137/160 Tholos-Bench scenarios against MiniCPM5-2B's 112/160 with Q4_K_M and llama.cpp JSON schema decoding on one Kaggle T4.
from datasets import load_dataset
data = load_dataset("mertkayacs/tholos-trajectories")
print(data["train"][0]["messages"])
The public dataset contains 263 trajectories: 237 training and 26 validation. Tholos-2B used a larger pool of 3,438 training and 349 validation runs. This subset alone does not reproduce that training run.
messages holds the system prompt, task, assistant steps and tool results. Assistant steps are JSON objects with thought, tool and args. Tool results arrive as user messages inside <tool_response>. meta records the scenario ID, category, template, domain, teacher, step count and token counts.
Qwen3.6-35B-A3B generated the public workspaces and agent steps. Runs were executed in the Tholos runtime, then checked against the final workspace state and lexical assertions. The examples cover table edits, notes, handoffs, approvals, conflicts, memory, follow-ups, web research and prompt injection. Generation pipeline.
Templates stay in one split; domains and names can appear in both. The public split differs from Tholos-2B's training split. Most public validation records were part of Tholos-2B's training, so use Tholos-Bench to evaluate that model.
All workspaces are invented, and all agent steps come from one teacher. Some prompts read awkwardly. Lexical checks can miss semantic errors. Injection examples deliberately contain hostile instructions; treat them as data and retain runtime approval rules.
Apache-2.0. The workspace text and agent steps are synthetic, generated by Qwen3.6-35B-A3B; the scenario templates and injection directives were authored for Tholos. The unpublished pool also includes GLM-5.2 and Kimi K3 outputs.
Tholos-2B weights and evaluation | GGUF files | Project.
An Eschatia Labs project. Mert Kaya.
Tholos trajectories: tool-use training data for small AI agents
0
7 commits
1 linked in READMEs
updated Oct 9, 2026
Executed English conversations for training small language models to call tools, follow approval rules and update a shared workspace. This public subset helped train Tholos-2B, which passes 137/160 Tholos-Bench scenarios against MiniCPM5-2B's 112/160 with Q4_K_M and llama.cpp JSON schema decoding on one Kaggle T4.
from datasets import load_dataset
data = load_dataset("mertkayacs/tholos-trajectories")
print(data["train"][0]["messages"])
The public dataset contains 263 trajectories: 237 training and 26 validation. Tholos-2B used a larger pool of 3,438 training and 349 validation runs. This subset alone does not reproduce that training run.
messages holds the system prompt, task, assistant steps and tool results. Assistant steps are JSON objects with thought, tool and args. Tool results arrive as user messages inside <tool_response>. meta records the scenario ID, category, template, domain, teacher, step count and token counts.
Qwen3.6-35B-A3B generated the public workspaces and agent steps. Runs were executed in the Tholos runtime, then checked against the final workspace state and lexical assertions. The examples cover table edits, notes, handoffs, approvals, conflicts, memory, follow-ups, web research and prompt injection. Generation pipeline.
Templates stay in one split; domains and names can appear in both. The public split differs from Tholos-2B's training split. Most public validation records were part of Tholos-2B's training, so use Tholos-Bench to evaluate that model.
All workspaces are invented, and all agent steps come from one teacher. Some prompts read awkwardly. Lexical checks can miss semantic errors. Injection examples deliberately contain hostile instructions; treat them as data and retain runtime approval rules.
Apache-2.0. The workspace text and agent steps are synthetic, generated by Qwen3.6-35B-A3B; the scenario templates and injection directives were authored for Tholos. The unpublished pool also includes GLM-5.2 and Kimi K3 outputs.
Tholos-2B weights and evaluation | GGUF files | Project.
An Eschatia Labs project. Mert Kaya.