quantcall-suite / README.md
happynood's picture
Add dataset card and smoke_v1 evaluation suite
06585d3 verified
|
Raw
History Blame Contribute Delete
1.84 kB
---
license: mit
language:
- en
tags:
- tool-use
- function-calling
- benchmark
- quantization
- evaluation
pretty_name: QuantCall Evaluation Suite
dataset_info:
features:
- name: id
dtype: string
- name: tier
dtype: string
- name: category
dtype: string
- name: query
dtype: string
- name: tools
sequence: string
- name: ground_truth_calls
sequence: string
- name: expects_call
dtype: bool
splits:
- name: smoke_v1
num_examples: 10
---
# QuantCall Evaluation Suite
Deterministic, versioned evaluation samples used by the
[QuantCall benchmark](https://github.com/Happynood/quant-toolcall-bench)
to measure how quantization degrades LLM function-calling reliability.
## Contents
| File | Description |
|------|-------------|
| `data/smoke_v1.jsonl` | T0 smoke tier — 10 hand-crafted instances, always available without a GPU |
| `data/schemas/tool_schemas.json` | Extracted JSON Schemas for all tools in `smoke_v1` |
## Format
Each instance in `smoke_v1.jsonl` is one JSON object per line:
```json
{
"id": "T0-001",
"tier": "T0",
"category": "simple",
"query": "What is the weather like in Paris?",
"tools": [{"name": "get_weather", "description": "...", "json_schema": {...}}],
"ground_truth_calls": [{"name": "get_weather", "arguments": {"city": "Paris"}}],
"expects_call": true
}
```
## Versioning
Files are version-pinned (`smoke_v1`, `smoke_v2`, …). Never overwrite a pinned
version; add new versions when the evaluation set changes. This ensures all
published results remain reproducible against the exact sample they were run on.
## Links
- GitHub: https://github.com/Happynood/quant-toolcall-bench
- Results dataset: https://huggingface.co/datasets/Happynood/quantcall-results
- Leaderboard: https://huggingface.co/spaces/Happynood/quantcall-leaderboard