Tholos-Bench: tool use by small local agent models

Tholos-Bench measures whether a model, acting as an agent in a shared workspace of tables, notes and tasks, leaves that workspace in the right state. It has 160 scenarios in 14 categories. Tholos-2B passes 137 of 160 against MiniCPM5-2B's 112, both as Q4_K_M files on one Kaggle T4 with llama.cpp JSON schema decoding.

Categories

Category	Scenarios	What the agent faces
approval	14	An action that needs the owner's approval first
ask	12	A gap only a person can fill, so the agent should ask
conflict	8	A row another writer changed after the agent read it
follow_up	8	Work that should be scheduled for later
handoff	14	A task that belongs to a teammate
injection	12	A page, table cell or note carrying instructions the agent should ignore
memory	10	Stored facts from earlier runs that should shape the work
notes	12	Writing, reading and appending to notes
nothing_to_do	8	A request where the right move is to change nothing
table_add	14	Adding rows
table_create	8	Creating a table
table_read_answer	12	Answering a question from table contents
table_update	14	Editing only the rows that match, leaving the rest
web_research	14	Reading local fixture pages and recording what they say
Total	160

Scoring

Each scenario sets up a workspace (agents, tools, tables, notes), sends one trigger (a task, a message or a schedule) and lets the model run for at most max_steps steps. Every model turn is one JSON step: a short thought, a tool name and its arguments. Tholos executes the call and returns the result.

Scoring reads the final state of the workspace and the run record. The expect list holds the checks, and a scenario passes only when every check holds. Checks include a row with given values, an unchanged table, a note containing a phrase, a task handed to the right teammate, an approval that was requested, a tool that was never called, and a finished run. The reference field is one passing sequence of tool calls; replaying it passes all 160 scenarios.

Web pages come from the fixtures field, so web research runs offline on .test hosts. respond holds the scripted answer to an approval or question, and interfere holds the edit another writer makes mid-run. The clock is pinned to a fixed UTC time, so prompts repeat exactly between runs.

Run it

Install Tholos (Python 3.11 or newer), download this dataset and point the runner at any OpenAI-compatible server:

uv tool install git+https://github.com/mertkayacs/tholos
hf download mertkayacs/tholos-bench --repo-type dataset --local-dir tholos-bench
llama-server -hf mertkayacs/Tholos-2B-GGUF:Q4_K_M --jinja -c 16384 -a tholos-2b --port 8080
tholos bench --base-url http://127.0.0.1:8080/v1 --model tholos-2b --scenarios tholos-bench/scenarios --out results.jsonl

pip install git+https://github.com/mertkayacs/tholos works too. The runner prints passed counts per category and writes one result per scenario, with the failed checks and the full transcript, to --out. --only <category> runs one category. The Tholos package ships the same 160 files, so --scenarios can be left out.

Load with datasets

import json
from datasets import load_dataset

bench = load_dataset("mertkayacs/tholos-bench", split="test")
row = bench[0]
expect = json.loads(row["expect"])

data/test.jsonl has one scenario per row. id, category, about, template, agent and max_steps are plain values. workspace, trigger, fixtures, respond, reference, expect and interfere are stored as JSON text, because their shape differs from scenario to scenario; json.loads gives back the original objects, and the three optional ones are null when a scenario lacks them. scenarios/ holds the same 160 files unchanged, one per scenario.

Results

Every row is a Q4_K_M file. Injection counts the 12 scenarios where a page, a table cell or a note carries instructions the agent should ignore.

Model, Q4_K_M	File	Passed, of 160	Injection, of 12
Qwen3.5-2B	1.28 GB	97	4
MiniCPM5-2B (base of Tholos-2B)	1.56 GB	112	8
Tholos-2B	1.56 GB	137	10
Llama 3.2 3B Instruct	2.02 GB	37	1
Granite 4.2-3B	2.24 GB	129	7
Qwen3.5-4B	2.74 GB	141	6

Tholos-2B passes 137, 25 more than MiniCPM5-2B, the model it was trained from. Qwen3.5-4B passes 141 with a file 1.75 times the size.

Every row ran on one Kaggle T4 with llama.cpp (CUDA build fdf5818), a 16,384-token context, JSON schema decoding, temperature 0, 512 tokens per step and the pinned clock. Only Tholos-2B was fine-tuned for the step format. LFM2.5-2.6B (1.67 GB) has no score, because it returned empty replies for every step under json_schema on this build.

Limits

Scores move with the server and decoding mode, so compare models inside one setup. The same two files on three setups:

Q4_K_M setup	Tholos-2B	MiniCPM5-2B
Kaggle T4, llama.cpp JSON schema	137	112
CPU, llama.cpp JSON schema	134	117
CPU, Ollama JSON mode with the GGUF repo template	136	96

CPU llama.cpp with the JSON schema is the reproducible path: on MiniCPM5-2B, a second run gave identical transcripts for all 160 scenarios. Tholos-2B was trained on Tholos tasks with the same thirteen tools, so this benchmark measures its target task and says nothing about other tool formats. The scenarios are constructed tasks, all in English. Twelve injection scenarios are too few to establish general safety.

Links

Tholos on GitHub | Tholos-2B | Tholos-2B-GGUF | Training trajectories | Project | Benchmark page

Citation

@misc{kaya2026tholosbench,
  author = {Kaya, Mert},
  title = {Tholos-Bench: Workspace Tasks for Small AI Assistants},
  publisher = {Eschatia Labs},
  year = {2026},
  url = {https://huggingface.co/datasets/mertkayacs/tholos-bench}
}

License

Apache-2.0, the license of the Tholos repository the scenarios come from.

An Eschatia Labs project. Mert Kaya.