---
license: mit
language:
- en
task_categories:
- question-answering
tags:
- finance
- tool-calling
- function-calling
- agent
- trajectory
- llm-evaluation
size_categories:
- n<1K
---

# FinTrace

**FinTrace** is a benchmark for evaluating LLM tool-calling (function-calling)
agents on financial queries, built on the Financial Modeling Prep (FMP) MCP
toolset. It contains expert-curated multi-turn tool-calling trajectories
across 30+ financial task categories, evaluated with a nine-metric rubric
spanning action correctness, execution efficiency, process quality, and
output quality.

Paper: *FinTrace* (COLM 2026) — link coming soon.

## Repository structure

| Path | Description | Status |
|---|---|---|
| `evaluation/testset.json` | 800-query evaluation set with output and golden trajectories | ✅ available |
| `trajectories/` | model-generated trajectories, one folder per model | 🔜 coming soon |
| `training/` | FinTrace-Training SFT / DPO preference data | 🔜 coming soon |

## Evaluation set (`evaluation/testset.json`)

800 entries. Each entry pairs a model **output trajectory** with a **golden
reference trajectory** (the best of three frontier-model runs, selected by an
LLM judge) for the same query:

| Field | Description |
|---|---|
| `id` | unique query id |
| `source_query` | the financial question |
| `task_type` / `task_type_bucket` | task category (32 types / 12 buckets) |
| `resource` / `data_source` | origin of the query |
| `difficulty_score` / `difficulty_tier` | difficulty annotation |
| `traj_len_bin` | trajectory length bin |
| `endpoints_called` | FMP endpoints invoked in the output trajectory |
| `reference_answer` | ground-truth answer |
| `output_answer` | final answer of the output trajectory |
| `reasoning` | reasoning summary of the output trajectory |
| `output_trajectory` | full multi-turn message list (reasoning, tool calls, tool responses) |
| `golden_trajectories` | golden reference trajectory for the same query |

## Usage

The trajectories are deeply nested, so we recommend downloading the raw JSON
directly rather than `load_dataset`:

```python
import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "YupengCao/FinTrace", "evaluation/testset.json", repo_type="dataset"
)
data = json.load(open(path))
print(len(data), data[0]["source_query"])
```

## Citation

Citation entry coming with the camera-ready release.