som_hermes_singleturn_nepali_v2 -->

A Nepali-language, single-turn function-calling dataset for supervised fine-tuning (SFT). It is derived from Somtharu181coder/som1_Hermes_function_calling_V1 through a translate → evaluate → filter pipeline. Only records that passed every automated check are published as the accepted set. Every rejected or failed record is accounted for in the audit summary below.

---

1. Dataset at a Glance

Item	Value
Dataset name	som_hermes_singleturn_nepali_v2
Source dataset	som1_Hermes_function_calling_V1
Target language	Nepali (Devanagari)
Format	Single-turn conversations with tool/function-calling structure
Records processed	927
Accepted (published)	529 (57.07%)
Not accepted	398 (42.93%)
of which: pipeline errors	22
of which: rejected by evaluation	376
Author	Som Narayan Tharu

Reconciliation: 529 accepted + 398 not accepted = 927 processed.

Known limitation of the audit

Some errored records have error: null, meaning the pipeline did not record why they failed. Example of the pattern (all processing fields are null):

{
  "id": "1a30feb9-c166-4394-91de-62b3da39cc34",
  "accepted": false,
  "errored": true,
  "reward": null,
  "answer_correct": null,
  "source_conversations": null,
  "translated_conversations": null,
  "raw_model_output": null,
  "source_provenance": {
    "category": "scheduling",
    "subcategory": "meeting_management",
    "task": "retrieve meeting details"
  },
  "error": null
}

For these records the root cause is not known and is reported as PIPELINE_ERROR_UNLABELED rather than guessed. Such records were excluded in full; none appear in the published data.

---

4. Data Format

Published fields (accepted set)

Field	Type	Description
id	string (UUID)	Unique record ID, stable across the audit and the published data.
source_conversations	list	Original conversation from the source dataset.
translated_conversations	list	Nepali translation. Tool schemas, function names and argument keys are kept in their original form.
reward	float	Final reward score from the evaluation stage.
answer_correct	bool	Result of the answer-correctness check.
source_provenance	object	category, subcategory, task of the original record.
translation_evaluation	object	Translation-quality evaluation output.
translation_semantic_evaluation	object	Semantic-fidelity evaluation output.

Audit-only fields (not in the published split)

accepted, errored, error, raw_model_output, input_format, result_metadata.

Example record

{
  "id": "<uuid>",
  "source_provenance": {"category": "<category>", "subcategory": "<subcategory>", "task": "<task>"},
  "source_conversations": [ ... ],
  "translated_conversations": [ ... ],
  "reward": 0.0,
  "answer_correct": true
}

Example

{"id": "2e0648a6-2023-417a-a5b9-f70304ff5d26", "conversations": [{"from": "system", "value": "You are a deep thinking AI, you may use extremely long chains of thought to deeply consider the problem and deliberate with yourself via systematic reasoning processes to help come to a correct solution prior to answering. You should enclose your thoughts and internal monologue inside tags, and then provide your solution or response to the problem.\n\nYou are a function calling AI model. You are provided with function signatures within XML tags. You may call one or more functions to assist with the user query. If available tools are not relevant in assisting with user query, just respond in natural conversational language. Don't make assumptions about what values to plug into functions. After calling & executing the functions, you will be provided with function results within XML tags. Here are the available tools:\n\n[\n {\n "name": "check_product_stock",\n "description": "Checks the current stock level for a given product ID.",\n "parameters": {\n "product_id": {\n "description": "The unique identifier of the product to check stock for.",\n "type": "string"\n }\n }\n }\n]\n\nFor each function call return a JSON object, with the following pydantic model json schema for each:\n{"title": "FunctionCall", "type": "object", "properties": {"name": {"title": "Name", "type": "string"}, "arguments": {"title": "Arguments", "type": "object"}}, "required": ["name", "arguments"]}\nEach function call should be enclosed within XML tags.\nExample:\n\n{"name": , "arguments": }\n"}, {"from": "human", "value": "उत्पादन XYZ123 को स्टक कति छ हेरिदिनु त?"}, {"from": "gpt", "value": "\n{"name": "check_product_stock", "arguments": {"product_id": "XYZ123"}}\n"}], "target_language": "ne-Deva", "source_query": "Can you check the stock level for product XYZ123?", "expected_tool_calls": [{"name": "check_product_stock", "arguments": {"product_id": "XYZ123"}}], "tools": [{"type": "function", "function": {"name": "check_product_stock", "description": "Checks the current stock level for a given product ID.", "parameters": {"type": "object", "properties": {"product_id": {"type": "string", "description": "The unique identifier of the product to check stock for."}}, "required": ["product_id"]}}}], "source_provenance": {"id": "2e0648a6-2023-417a-a5b9-f70304ff5d26", "category": "eCommerce", "subcategory": "Inventory Management", "task": "Check product stock level"}, "localizer_model": "google:gemini-3.5-flash-lite", "policy_model": "google:gemini-3.5-flash-lite"}

Intended Use

• SFT / instruction tuning of models for Nepali function calling and tool use.
• Evaluating Nepali tool-use behavior of existing models.
• Research on low-resource-language function calling.

Limitations and Biases

• Single-turn only. No multi-turn dialogue, no tool-response chaining.

• Machine-translated. Passed automated evaluation, but was not fully human-verified. Residual translation errors and unnatural phrasing are possible.

• Evaluator bias. Acceptance depends on automated evaluators, which have their own errors, so some good records may be rejected and some weak ones accepted.

• Category skew. The accepted set reflects survivorship per category (see 3.2); under-represented categories may have lower coverage.

• Inherited issues. Any errors or biases in the source dataset carry over.

  License and Attribution

Released under apache-2.0. This dataset is derived from som1_Hermes_function_calling_V1, which in turn derives from Hermes function-calling data. Check and comply with the upstream license and attribution terms before reuse.

Citation

@dataset{tharu_som_hermes_singleturn_nepali_v2,
  author    = {Tharu, Som Narayan},
  title     = {som\_hermes\_singleturn\_nepali\_v2: A Nepali Single-Turn Function-Calling Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Somtharu181coder/som_hermes_singleturn_nepali_v2}
}

Contact

Som Narayan Tharu, Hugging Face: Somtharu181coder