File size: 2,352 Bytes
3e7b2e5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | ---
license: apache-2.0
language:
- en
tags:
- genomics
- tool-calling
- function-calling
- qwen
- mlx
pretty_name: CodeXomics-ToolCalling-v1 (V5 training data)
---
# CodeXomics-ToolCalling-v1
This dataset contains the supervised tool-calling trajectories used to fine-tune
**CodeXomics-ToolAgent-4B-v1** (internally `qwen3.5:4b-codexomics-tools-v5`) for the CodeXomics
genomics workbench (an AI-native genome browser; source:
[github.com/Scilence2022/CodeXomics](https://github.com/Scilence2022/CodeXomics)). It is released
as the reproducibility artifact for the corresponding paper section and is licensed under
Apache-2.0.
## Contents
- `train.jsonl` — 373 supervised examples
- `valid.jsonl` — 123 validation examples
- `test.jsonl` — 30 holdout examples
- `manifest.json` — dataset schema, eligibility metadata, registry hash, and leakage-control record
- `token-stats.json` — render-length statistics per split (3,072-token budget)
Each line is an MLX-LM tool-calling example: a system prompt, a user instruction, and assistant
`tool_calls` in the OpenAI function-calling shape. Multi-step examples include the returned tool
results and use the cross-round reference syntax `{tool_name.path}`.
## Composition
| Split | Single-call | Multi-step | Variants | Synthetic chains | Total |
| --- | ---: | ---: | ---: | ---: | ---: |
| Train | 85 | 18 | 10 | 260 | 373 |
| Validation | 66 | 8 | 4 | 45 | 123 |
| Test | 21 | 8 | 1 | 0 | 30 |
Source trajectories (678 records) were admitted only after passing: (1) strong-model semantic
replay (DeepSeek V4 Flash), (2) fixture executability, (3) state verification for stateful
records, and (4) benchmark leakage control. All 172 CodeXomics Benchmark tests are fingerprinted
and excluded (`benchmark_scope.use: leakage-gate-only`).
## Training schema
Six candidate tool schemas per request (gold tool always included), function descriptions truncated
to 160 characters, parameter defaults preserved, semantically hard distractors, and a
deterministic-randomized gold position.
## Usage
```bash
huggingface-cli download <this-dataset> --local-dir codexomics-v5-data
```
## Citation
```bibtex
@software{codexomics-toolagent-v1,
title = {CodeXomics-ToolCalling-v1},
author = {Song, Lifu},
year = {2026},
license = {Apache-2.0},
publisher = {Hugging Face}
}
```
|