---
configs:
  - config_name: default
    data_files:
      - split: ne_deva
        path: data/ne_deva.jsonl
      - split: ne_latn
        path: data/ne_latn.jsonl
license: apache-2.0
language:
- ne
- en
task_categories:
- text-generation
tags:
- function-calling
- tool-use
- nepali
- sharegpt
- hermes
- synthetic
size_categories:
- 1K<n<10K
---

# hermes-function-calling-nepali

Single-turn function calling with the user request re-spoken in Nepali — Devanagari
(`ne_deva`) and romanized Latin (`ne_latn`) — voice-assistant style, with tool calls
verified against the English ground truth. Tool schemas and expected calls are unchanged
from [NousResearch/hermes-function-calling-v1](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1)
(`func_calling_singleturn`); only the user turn was localized.

Generated with [HimalayaAI/gymkhana](https://github.com/HimalayaAI/gymkhana)'s
`multilingual-tool-use` environment:

1. **Localizer** (`gpt-5.6-luna-pro`) rewrites the English request as a Nepali speaker
   would say it to a voice assistant. A deterministic gate rejects rewrites that drop an
   identifier-like argument value, URL or e-mail the tool needs, or that are in the wrong
   script.
2. **Policy** (`deepseek-v4-flash-0731`) gets the English tool schemas via native tool
   calling plus the Nepali request; its provider-returned reasoning is captured.
3. **Verifier**: calls must match the English ground truth exactly (name and arguments,
   all calls present, order-independent). A row is exported only if one of 4 rollouts
   matched; that rollout is the one exported.

Both models served via Nous Portal.

## Splits

| split | script | rows |
|---|---|---|
| `ne_deva` | Nepali, Devanagari | 584 |
| `ne_latn` | Nepali, romanized Latin | 602 |

543 source rows appear in both splits — same request and ground truth in two scripts,
useful for measuring script transfer.

## Columns

- `id`
- `conversations`
- `tools`
- `category`
- `subcategory`
- `task`
- `target_language`
- `source_query`
- `expected_tool_calls`
- `localizer_model`
- `policy_model`

`conversations` is ShareGPT in Hermes format: `system` carries the tool schemas inside
`<tools>`, `human` is the Nepali request, `gpt` carries reasoning inside `<think>…</think>`
followed by `<tool_call>…</tool_call>` blocks. `tools` and `expected_tool_calls` are JSON
strings.

## Example (`ne_latn`)

```
human: Mero project ko naya feature review ko lagi ready cha. Pahila feature_branch lai
       origin remote repository ma push gardinu, ani feature_branch bata main branch ko
       lagi pull request banaidinu.
gpt:   <think>The user wants two things: push the branch, then open a PR …</think>
       <tool_call>{"name": "push_branch_to_remote", "arguments": {"branch_name": "feature_branch", "remote_name": "origin"}}</tool_call>
       <tool_call>{"name": "create_pull_request", "arguments": {...}}</tool_call>
```

## Stats

- rows: 1186 (584 + 602)
- source rows attempted per script: 1090 (first 1100 of `func_calling_singleturn`)
- excluded: localization gate (dropped values or wrong script), and rows where no rollout
  of 4 reproduced the ground-truth calls

## License

Apache-2.0, inherited from the source dataset.