Somtharu181coder/number_of_death_by_sex_hermes_calling

Nepal Education Enrollment Statistics – Hermes Function-Calling Dataset

1. Overview

This dataset contains 16,760 single-turn function-calling records in Hermes / ShareGPT conversation format. Each record pairs a natural-language request for education enrollment statistics with the exact tool call that satisfies it. Tool-call arguments are grounded in the administrative hierarchy of an Excel source workbook (Annex 4 – Enrollment Details): every province, district, and local-level value appearing in a request is reproduced verbatim in the corresponding tool call.

The dataset is designed for the Gymkhana multilingual tool-use environment and supports supervised fine-tuning and evaluation of tool-calling behavior, in particular faithful argument copying and metric selection.

Scope note. The dataset contains tool requests only. It does not contain tool responses or enrollment figures.

2. Dataset at a Glance

Property	Value
File	education_numberofdeathysex_hermes.jsonl
Format	JSON Lines, UTF-8, LF line endings, no BOM
Records	16,760
File size	51,870,984 bytes (≈ 49.5 MiB); ≈ 3.1 KB per record
Split	train (single split)
Category / Subcategory / Task	Education / Enrollment Statistics / Education Statistics
Target-language tag	ne-Deva (all records)
Tool schema	One function: get_education_statistics
Turns per record	3 (system → human → gpt)
Tool calls per record	Exactly 1
Generator	excel_to_hermes_grounded.py

3. Data Structure

3.1 Field Reference

Field	Type	Description
id	string	Deterministic UUID (version 5); unique across the file.
conversations	list[object]	Ordered turns; each has from (system, human, gpt) and value.
tools	string	JSON-serialized tool schema (compact). Identical in every record.
category	string	Top-level domain (Education).
subcategory	string	Enrollment Statistics.
task	string	Education Statistics.
target_language	string	Language/script tag for the multilingual environment (ne-Deva).
source_query	string	Source-language request text; identical to the human turn.
expected_tool_calls	string	JSON-serialized list holding the reference tool call.

3.2 Tool Schema

name: get_education_statistics
description: >-
  Retrieve education statistics for a specified province, district, and local level.
  The metrics field specifies which education statistics are requested.
parameters:
  province:    { type: string, required: true }
  district:    { type: string, required: true }
  local_level: { type: string, required: true }
  metrics:
    type: array
    minItems: 1
    items: { type: string, enum: [a_male, b_female, rowtotal] }
    required: true
additionalProperties: false

Metric ID	Meaning
a_male	Male enrollment
b_female	Female enrollment
rowtotal	Total enrollment

3.3 Request Template

Every human turn follows a single template:

Please provide the {metric} statistics for local level [{local_level}], district [{district}], and province [{province}]. Keep all values inside [] exactly unchanged.

The gpt turn is a single Hermes tool call in compact JSON:

<tool_call>
{"name":"get_education_statistics","arguments":{"province":"…","district":"…","local_level":"…","metrics":["…"]}}
</tool_call>

3.4 Example Record

{
  "id": "50b80e24-392c-55bd-9a08-48b120c3b513",
  "conversations": [
    {"from": "system", "value": "You are a function calling AI model. … <tools>\n[…]\n</tools>\nWhen you decide to call a function, output the function call inside <tool_call> and </tool_call> tags as valid JSON."},
    {"from": "human",  "value": "Please provide the b_female statistics for local level [NEPAL], district [NEPAL], and province [NEPAL]. Keep all values inside [] exactly unchanged."},
    {"from": "gpt",    "value": "<tool_call>\n{\"name\":\"get_education_statistics\",\"arguments\":{\"province\":\"NEPAL\",\"district\":\"NEPAL\",\"local_level\":\"NEPAL\",\"metrics\":[\"b_female\"]}}\n</tool_call>"}
  ],
  "tools": "[{\"type\":\"function\",\"function\":{\"name\":\"get_education_statistics\", …}}]",
  "category": "Education",
  "subcategory": "Enrollment Statistics",
  "task": "Education Statistics",
  "target_language": "ne-Deva",
  "source_query": "Please provide the b_female statistics for local level [NEPAL], district [NEPAL], and province [NEPAL]. Keep all values inside [] exactly unchanged.",
  "expected_tool_calls": "[{\"name\":\"get_education_statistics\",\"arguments\":{…}}]"
}

4. Static Audit Report

4.1 Methodology

The audit is static: it inspects the released file only, without executing any tool or model. All 16,760 lines were parsed with a strict JSON parser, and each record was checked for schema conformity, internal consistency between fields, argument grounding, tool-call validity, serialization hygiene, and content uniqueness. Results below are exact counts, not samples.

4.2 Summary of Results

#	Check	Result	Status
1	Lines parsed as valid JSON (blank or malformed lines)	16,760 / 16,760 (0)	Pass
2	Uniform key set across all records	9 keys, 100 %	Pass
3	id uniqueness	16,760 unique; all UUID v5	Pass
4	Turn structure system → human → gpt	100 %	Pass
5	Tool schema identical across records	1 distinct schema	Pass
6	Embedded <tools> block in system turn equals tools field	0 mismatches	Pass
7	gpt turn format (<tool_call> wrapper, compact JSON)	100 % conformant	Pass
8	gpt tool call equals expected_tool_calls	0 mismatches	Pass
9	Exactly one tool call per record	100 %	Pass
10	source_query equals human turn	0 mismatches	Pass
11	Bracketed values in query equal tool-call arguments (local level, district, province)	16,760 / 16,760	Pass
12	Requested metric in query equals metrics argument	0 mismatches	Pass
13	Argument key order (province, district, local_level, metrics)	100 %	Pass
14	Leading/trailing whitespace in any turn	0	Pass
15	Encoding hygiene (LF endings, no BOM, trailing newline)	Clean	Pass
16	Distinct records by conversation content	2,512 of 16,760	Finding F1
17	Script of human turn vs. target_language tag	0 Devanagari characters; all ASCII	Finding F2
18	Naming consistency of administrative units	Mixed casing/format variants	Finding F3

4.3 Distribution Profile

Metric balance. The three metrics are close to evenly distributed.

Metric	Records	Share
b_female	5,651	33.72 %
a_male	5,572	33.25 %
rowtotal	5,537	33.04 %

Geographic hierarchy. The file covers 838 distinct location tuples, each contributing exactly 20 records.

Level	Location tuples	Records	Share of records
National aggregate (NEPAL / NEPAL / NEPAL)	1	20	0.12 %
Province aggregates (province name in upper case, repeated at all three levels)	7	140	0.84 %
District aggregates (local level ends with (TOTAL))	77	1,540	9.19 %
Local units (rural municipalities, municipalities, sub-metropolitan and metropolitan cities)	753	15,060	89.86 %
Total	838	16,760	100 %

The 753 local units correspond to 736 units named … Gaunpalika / … Municipality and 17 metropolitan or sub-metropolitan cities.

Text length. human turns range from 146 to 216 characters (median 168). The system prompt is identical in all records.

4.4 Findings

ID	Severity	Finding	Impact	Recommendation
F1	Medium	Content-level duplication. Only 2,512 distinct conversations exist across 16,760 records (one per location–metric pair present). 14,248 records (85.02 %) repeat another record's content and differ only in id. Group sizes range from 1 to 16 (most commonly 5–8).	Inflates dataset size, over-weights repeated pairs during training, and risks train/eval leakage if the file is split by record.	De-duplicate on the conversations content before splitting, or split by location-metric key. Retain duplicates only if intentional up-sampling is desired.
F2	Medium	Language tag versus content. All records carry target_language = ne-Deva, yet every human and gpt value is ASCII English/Latin text; no Devanagari characters are present. source_query is identical to the human turn, indicating no translation has been applied in this file.	Consumers that assume Nepali-script user turns will receive English text.	Confirm whether the Gymkhana environment translates at runtime using target_language. If the file is meant to be pre-translated, translate the request text while leaving bracketed values untouched.
F3	Low	Naming variants inherited from the source. Province appears as Province-1 in local rows but PROVINCE - 1 in aggregate rows; other provinces appear in both title case and upper case; district aggregates use upper-case district names (e.g., LALITPUR) while local rows use title case. Source spellings such as Metropolitian are retained.	Case-sensitive joins across levels will not match. The verbatim form is intentional (see §5.3).	Normalize only in downstream analytics, never in the grounded arguments.
F4	Low	Partial metric coverage for two locations. Molung Gaunpalika (Province-1 / Okhaldhunga) has no a_male record; Gaidahawa Gaunpalika (Lumbini / Rupandehi) has no rowtotal record. All other 836 locations cover all three metrics.	Minor evaluation imbalance.	Add the missing pairs if full coverage is required.
F5	Informational	Redundant tool-schema serialization. The schema is stored in both the system turn and the tools field in every record.	Adds size overhead; supports both training and evaluation consumers.	Retain unless storage is a constraint.
F6	Informational	Single-turn, request-only design. No tool-response or final-answer turns are present, and no statistical values appear in the file.	Suitable for tool-call generation; unsuitable for response-synthesis training.	None required; noted for scope clarity.

4.5 Machine-Readable Audit Manifest

audit:
  file: education_numberofdeathysex_hermes.jsonl
  type: static
  bytes: 51870984
  records: 16760
  parse_failures: 0
  schema:
    keys: [id, conversations, tools, category, subcategory, task, target_language, source_query, expected_tool_calls]
    uniform_keyset: true
    turn_pattern: [system, human, gpt]
  identifiers:
    unique_ids: 16760
    uuid_version: 5
  tools:
    distinct_schemas: 1
    function: get_education_statistics
    system_block_matches_tools_field: true
  tool_calls:
    per_record: 1
    gpt_equals_expected: true
    grounding_mismatches: 0
    metric_mismatches: 0
  metrics:
    b_female: 5651
    a_male: 5572
    rowtotal: 5537
  locations:
    distinct_tuples: 838
    records_per_tuple: 20
    national: 1
    province_aggregates: 7
    district_aggregates: 77
    local_units: 753
  uniqueness:
    distinct_conversations: 2512
    redundant_records: 14248
    redundancy_ratio: 0.8502
  language:
    target_language_tag: ne-Deva
    devanagari_characters_in_text: 0
  hygiene:
    line_endings: LF
    bom: false
    trailing_newline: true
    whitespace_padded_turns: 0
  findings: [F1, F2, F3, F4, F5, F6]

4.6 Reproducing the Core Checks

import json, re, collections as C

rows = [json.loads(l) for l in open("education_numberofdeathysex_hermes.jsonl", encoding="utf-8")]
assert len({r["id"] for r in rows}) == len(rows)

grounded = 0
for r in rows:
    q = r["conversations"][1]["value"]
    args = json.loads(r["expected_tool_calls"])[0]["arguments"]
    brackets = re.findall(r"\[([^\]]*)\]", q)
    grounded += brackets[:3] == [args["local_level"], args["district"], args["province"]]

distinct = len({json.dumps(r["conversations"], sort_keys=True, ensure_ascii=False) for r in rows})
print(f"grounded={grounded}/{len(rows)}  distinct_conversations={distinct}")

5. Data Construction and Processing

The processing applied by excel_to_hermes_grounded.py is minor and structural. The steps below are those verifiable from the released file.

5.1 Source and Record Construction

1. Source. Administrative units (national, province, district, local level) and the three enrollment metrics are taken from the Annex 4 – Enrollment Details workbook.
2. Templating. Each record is built from one fixed request template (§3.3), populated with a location tuple and a single metric.
3. Tool-call synthesis. The reference tool call is generated directly from the same tuple and metric, so query and call cannot diverge. The call is stored twice: as the gpt turn (Hermes <tool_call> block) and as expected_tool_calls (JSON string).
4. Schema embedding. The tool schema is embedded in the system turn between <tools> tags and in the tools field.
5. Metadata. category, subcategory, task, and target_language are attached as constant fields; source_query mirrors the human turn.

5.2 Serialization Conventions

Aspect	Convention
Encoding	UTF-8, LF, one JSON object per line, trailing newline
Tool-call JSON	Compact separators (, and :), fixed key order
Nested JSON fields	tools and expected_tool_calls stored as JSON strings
Identifiers	UUID version 5 (deterministic, namespace-based)
Role labels	ShareGPT-style: system, human, gpt

5.3 Value Preservation

• Location names are copied verbatim from the workbook, including capitalization, hyphenation, parenthetical (TOTAL) markers, and source spellings.
• The request explicitly instructs the model to keep bracketed values unchanged, and the audit confirms 100 % agreement between bracketed values and tool-call arguments.
• No case-folding, stripping, transliteration, or spelling correction is applied to grounded values.

5.4 Sampling Structure

Each of the 838 location tuples contributes exactly 20 records, with the metric varying across those records. As a consequence, location–metric pairs recur several times (see Finding F1).

6. Intended Use and Limitations

Intended use

• Supervised fine-tuning of models for schema-conformant function calling.
• Evaluation of argument fidelity (verbatim copying of bracketed values) and metric selection.
• Testing multilingual tool-use pipelines within the Gymkhana environment.

Limitations

• Single tool, single turn, and a single request template; linguistic diversity is intentionally low.
• No tool responses, numerical statistics, or answer-synthesis turns.
• Substantial duplication (Finding F1) and an unresolved language tag (Finding F2) should be addressed before use in benchmarking or in train/validation splitting.
• Not suitable as a source of enrollment statistics.

7. Licensing and Attribution

The dataset is derived from the Annex 4 – Enrollment Details workbook. The license field is set to other pending confirmation of the source workbook's terms; users should verify the applicable terms before redistribution.

8. Citation

@dataset{nepal_education_enrollment_hermes,
  title  = {Nepal Education Enrollment Statistics -- Hermes Function-Calling Dataset},
  year   = {2026},
  note   = {Generated with excel_to_hermes_grounded.py for the Gymkhana multilingual tool-use environment}
}