README.md

---
license: apache-2.0
language:
- en
- es
task_categories:
- text-generation
pretty_name: Network Agent (FunctionGemma) EN/ES Tool-Calling Dataset
tags:
- synthetic
- forjal
- function-calling
- functiongemma
- tool-use
- two-turn
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
  - split: validation
    path: data/validation.jsonl
  - split: test
    path: data/test.jsonl
- config_name: en
  data_files:
  - split: train
    path: data/en/train.jsonl
  - split: validation
    path: data/en/validation.jsonl
  - split: test
    path: data/en/test.jsonl
- config_name: es
  data_files:
  - split: train
    path: data/es/train.jsonl
  - split: validation
    path: data/es/validation.jsonl
  - split: test
    path: data/es/test.jsonl
---

# Network Agent (FunctionGemma) EN/ES Tool-Calling Dataset

Synthetic two-turn tool-calling dataset for a small on-device network assistant, exported in the
`functiongemma_tool_calling` format and compatible with the Google/Hugging Face FunctionGemma flow
(`apply_chat_template(messages, tools=tools)`).

The agent answers user questions about the home network by calling one of three tools on the first
turn and then producing a short natural-language answer on the second turn, using the tool result
as fact.

## Configurations

- `default` — combined bilingual English + Spanish (1112 rows).
- `en` — English only (556 rows).
- `es` — Spanish only (556 rows).

| Config | train | validation | test | total |
|--------|------:|-----------:|-----:|------:|
| default (en+es) | 842 | 118 | 152 | 1112 |
| en | 421 | 59 | 76 | 556 |
| es | 421 | 59 | 76 | 556 |

## Tools

- `get_number_connected_users` — number of devices connected to the network.
- `check_internet_status` — `0` = working, `1` or greater = failure.
- `get_internet_speed` — current internet speed in Mbps.

All tools take an empty argument object.

## Row format

Each row is a chat example with `messages` and `tools`:

- `messages[0]` — `developer` prompt (English or Spanish variant).
- First-turn rows: `developer, user, assistant(tool_calls)`.
- Second-turn rows: `developer, user, assistant(tool_calls), tool, assistant(content)`.
- Direct-response rows (small talk / out-of-scope refusal): `developer, user, assistant(content)`.

`role: "developer"` is intentional for FunctionGemma and should not be rewritten to `system`.

## Coverage modules

`first_turn_core`, `second_turn_core_success`, `second_turn_error`, `first_turn_edge`,
`mixed_robustness`, `second_turn_conditional`, plus small-talk and out-of-scope refusal rows.
See `reports/` for per-split composition and balance reports.

## Provenance

Derived and rebalanced from `Forjal/forjal-network-agent-v4-9847d936`. This release normalizes
second-turn rows to a single final assistant message, replaces malformed first-turn text with the
tool call, balances tools across splits, and adds a Spanish translation so the model can respond in
English or Spanish based on the user's language.