Datasetpublic

sabin1234/drug_poisoning_nepali_sharegpt

README — drug_poisoning_nepali_sharegpt_cleaned.jsonl This README file provides detailed information about the drug_poisoning_nepali_sharegpt_cleaned.jsonl dataset — file format, schema, source, subject matter, question pattern diversity, answer behaviour diversity, geographic/temporal coverage, and statistical analysis, all presented in tables. 1. General File Information Detail Value File name drug_poisoning_nepali_sharegpt_cleaned.jsonl Format… See the full description on the dataset page: https://huggingface.co/datasets/sabin1234/drug_poisoning_nepali_sharegpt.

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face

0likes47downloads

Dataset Card

README — `drug_poisoning_nepali_sharegpt_cleaned.jsonl`

This README file provides detailed information about the drug_poisoning_nepali_sharegpt_cleaned.jsonl dataset — file format, schema, source, subject matter, question pattern diversity, answer behaviour diversity, geographic/temporal coverage, and statistical analysis, all presented in tables.

1. General File Information

Detail	Value
File name	`drug_poisoning_nepali_sharegpt_cleaned.jsonl`
Format	JSON Lines (`.jsonl`) — each line is an independent JSON object
Total records (rows)	56,448
File size	~77 MB
Conversation structure	ShareGPT-style (`human` and `gpt` roles inside a `conversations` array)
Number of turns per record	2 (1 human + 1 gpt) — 100% single-turn
Language	Nepali (ne / npi)
Script	Devanagari (Deva)
Duplicate records	0 — all `id`s, questions, and answers are completely unique

2. Dataset Domain Overview

Aspect	Detail
Original source data	Centers for Disease Control and Prevention (CDC), USA
Domain	County-level population and "Drug Poisoning Mortality" (age-adjusted death rate)
Variables measured	(1) Annual population of the relevant county/district, (2) Age-adjusted mortality rate category (range)
Geographic unit	US State + County/District
Time range	1999 to 2016 (18 years)
Data type	Synthetic — question-answer pairs generated in Nepali by a model, based on real CDC statistics
Generating model	`himalaya-gemma-4-e2b-it`
Task type	Instruction-following (single-fact retrieval / lookup QA)

3. JSON Schema (fields in each record)

Field name	Data type	Description	Example value
`id`	string	Unique record identifier	`drug_poisoning_sft_000001`
`conversations`	array	Array containing 2 objects with `human` and `gpt` roles	see below
`conversations[0].from`	string	Always `human`	`human`
`conversations[0].value`	string	User's question (in Nepali)	—
`conversations[1].from`	string	Always `gpt`	`gpt`
`conversations[1].value`	string	Model's answer (in Nepali)	—
`source`	string	Type of data origin	`synthetic`
`source_name`	string	Original data-providing organization	`Centers for Disease Control and Prevention`
`source_repo`	string	Source repository name	`local_generation`
`source_config`	string	Dataset configuration/domain name	`drug_poisoning_mortality`
`source_split`	string	Data split	`train`
`source_revision`	string	Version number	`v1.0`
`source_row_id`	string	Row sequence number of the original table	`1`, `2`, `3` ...
`language`	string	Language code (short)	`ne`
`language_code`	string	Language code (ISO)	`npi`
`script`	string	Script	`Deva`
`license`	string	License detail	`Public Domain - U.S. Government`
`license_tier`	string	License class	`permissive`
`task_type`	string	Task type	`instruction-following`
`generation_type`	string	Generation method	`synthetic`
`condition`	string	Data condition	`model-generated`
`url`	string	Source URL (empty here)	`""`
`metadata_json`	string (JSON-encoded)	Additional metadata (`domain`, `model`, `generation_number`)	`{"domain":"drug_poisoning_mortality","model":"gemini-3.5-flash-lite","generation_number":1}`

4. Value Diversity of Metadata Fields (Constant vs Variable Fields)

Field	Number of unique values	Status
`id`	56,448 (all different)	Fully variable
`source`	1 (`synthetic`)	Constant
`source_name`	1 (`Centers for Disease Control and Prevention`)	Constant
`source_repo`	1 (`local_generation`)	Constant
`source_config`	1 (`drug_poisoning_mortality`)	Constant
`source_split`	1 (`train`)	Constant
`source_revision`	1 (`v1.0`)	Constant
`source_row_id`	56,448 (all different, sequential)	Variable
`language`	1 (`ne`)	Constant
`language_code`	1 (`npi`)	Constant
`script`	1 (`Deva`)	Constant
`license`	1 (`Public Domain - U.S. Government`)	Constant
`license_tier`	1 (`permissive`)	Constant
`task_type`	1 (`instruction-following`)	Constant
`generation_type`	1 (`synthetic`)	Constant
`condition`	1 (`model-generated`)	Constant
`metadata_json → model`	1 (`himalaya-gemma-4-e2b-it`)	Constant
`metadata_json → domain`	1 (`drug_poisoning_mortality`)	Constant
`metadata_json → generation_number`	56,448 (all different)	Variable

5. Basic Statistics

Metric	Human (question)	GPT (answer)
Total count	56,448	56,448
Unique text count	56,448 (100%)	56,448 (100%)
Average character length	132.8	140.1
Minimum character length	117	108
Maximum character length	180	188
Average word count	20.4	20.8
Minimum word count	17	17
Maximum word count	29	30

6. Question Pattern Diversity

All questions in the dataset are surface-level distinct (56,448 out of 56,448 are unique), but their underlying structure (template) is based on a limited number of paraphrase variants. Even after normalizing (removing numbers like year, population, ID), many different sentence structures are found, whose main characteristics are shown in the tables below.

6.1 Question Ending Patterns — 19 distinct patterns total

No.	Ending pattern	Count	Percentage
1	...what was the category?	9,167	16.24%
2	...which category was it in?	4,763	8.44%
3	...which range was it in?	4,714	8.35%
4	...which category is it in?	4,537	8.04%
5	Identify the range.	2,415	4.28%
6	...which range does it show?	2,402	4.26%
7	...which class did it fall into?	2,385	4.23%
8	...what range was it placed in?	2,375	4.21%
9	...what does the record show?	2,355	4.17%
10	Also state the range.	2,346	4.16%
11	...which range did it fall into?	2,346	4.16%
12	...what was the population?	2,338	4.14%
13	State the rate range.	2,333	4.13%
14	...what was the classification?	2,331	4.13%
15	...which level was it at?	2,305	4.08%
16	...what category was recorded?	2,204	3.90%
17	...which range were they in?	2,202	3.90%
18	...which category is shown?	2,192	3.88%
19	...mentioned in the statistics.	738	1.31%

6.2 Diversity by Question Sentence Type (Sentence Mood)

Sentence type	Count	Percentage	Characteristic
Interrogative (ends in `?`)	48,616	86.11%	Questions like "...how much?", "...which category was it in?"
Imperative/Request (ends in `।`)	7,832	13.89%	Requests like "...state it", "...identify it"

6.3 Diversity by Question Opening Structure

Opening type	Count	Percentage	Example
Year/time-first	24,084	42.67%	"In the year 2011...", "For the year 1999..."
State/district-first	32,364	57.33%	"In the state of Texas...", "Autauga district..."

6.4 Information Pattern within Questions

Fact-requesting pattern	Present in how many questions	Percentage
Population asked	56,448	100%
Mortality rate category asked	56,448	100%
Both facts asked together (compound question)	56,448	100%
Comparison/trend question	0	0%

Note: Every question in the dataset follows the same "single-fact lookup" pattern — asking for the population and mortality rate category of a specific state + district + year. There are no complex questions such as comparisons, trends, or multi-year queries.

6.5 Synonym / Lexical Variation

Different Nepali words used to express the same concept (paraphrase diversity):

Concept	Different words/phrases used
"category/level"	श्रेणी, वर्ग, दायरा, तह, वर्गीकरण (category, class, range, level, classification)
"year"	वर्ष, साल, समय (year, year, time)
"statistics/record"	तथ्याङ्क, अभिलेख, डेटा, विवरण (statistics, record, data, detail)
"asking verb"	कति छ, के थियो, बताउनुहोस्, पहिचान गर्नुहोस्, देखाउँछ, उल्लेख छ (how much is it, what was it, state it, identify it, shows, is mentioned)
"was/is" (tense)	Both past tense (थियो, परेको थियो) and present tense (छ, देखाउँछ) are used

7. Answer Behaviour Diversity

7.1 Answer Opening Patterns (Top 20 — out of 3,330 distinct openings total)

Opening phrase	Count	Percentage
Based on the record...	2,271	4.02%
In that year...	2,261	4.01%
In that year...	2,186	3.87%
In 2010...	543	0.96%
In 1999...	521	0.92%
In 2009...	519	0.92%
In 2006...	518	0.92%
In 2013...	515	0.91%
In 2002...	513	0.91%
In 2003...	512	0.91%
In 2001...	500	0.89%
In 2005...	499	0.88%
In 2016...	497	0.88%
In 2012...	494	0.87%
In 2015...	490	0.87%
In 2007...	482	0.85%
In 2014...	478	0.85%
In 2008...	475	0.84%
In 2011...	461	0.82%
In 2004...	456	0.81%

7.2 Answer Content Structure

Structural element	Description	Presence
Population figure	Always a concrete number (e.g. 42963)	100% of answers
Mortality rate category	Always a numeric range (e.g. 2-3.9) or a bound like `<2`	100% of answers
Sentences per answer	Usually 2 sentences (one for population, one for mortality rate); sometimes both combined in 1 sentence	Variable
Past/present tense usage	Answers match the tense of the question, using both past and present	Mixed
Filler/Framing phrases	Various introducers such as "based on the record", "the statistics show", "according to the details"	High diversity

7.3 Distribution of Mortality Rate Categories

Category (rate per 100,000 population)	Estimated count (based on observed samples)
< 2	Present in high numbers
2 – 3.9	High
4 – 5.9	Most common overall
6 – 7.9	High
8 – 9.9	Moderate-high
10 – 11.9	Moderate
12 – 13.9	Moderate
14 – 15.9	Low
16 – 17.9	Low
18 – 19.9	Rare
20 – 21.9	Rare
22 – 23.9	Rare
24 – 25.9	Rare
26 – 27.9	Very rare
28 – 29.9	Very rare

These categories resemble CDC's standard "age-adjusted death rate range/category" format — where lower-numbered ranges (such as <2 or 2-3.9) are more common and higher ranges (up to 28-29.9) appear only in rare counties.

8. Geographic Coverage

8.1 Top 20 States by Presence

No.	State	Mentions in questions (approx.)
1	Texas	1,559
2	North Carolina	964
3	Georgia	961
4	Virginia	889
5	Kentucky	746
6	South Dakota	739
7	Missouri	733
8	Illinois	653
9	Kansas	657
10	Iowa	611
11	West Virginia	570
12	Tennessee	561
13	Nebraska	554
14	Indiana	544
15	Mississippi	533
16	Minnesota	530
17	Ohio	524
18	North Dakota	518
19	Michigan	517
20	Arkansas	498

Overall geographic statistic	Value
Total distinct states found	30+ (abbreviated US state names)
Total distinct county/district names (word-based estimate)	1,650+
Geographic level	US State + County level

9. Temporal Coverage

Metric	Value
Years used	1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016
Total distinct years	18
Most frequently appearing years (example)	2007, 2012, 2015 (each ~3,138 times)
Time range type	Yearly snapshot; no trend/multi-year questions

10. Sample Records (showing diverse styles)

ID	Question style	Sample question	Sample answer
drugpoisoningsft_000001	State + District + Year, interrogative, both facts together	What was the population of Autauga district in the state of Alabama in the year 1999, and which category is the age-adjusted mortality rate from drug poisoning in?	The population of Autauga district in 1999 was 42963. The age-adjusted mortality rate category from drug poisoning was 2-3.9.
drugpoisoningsft_015001	Year-first + imperative	For the year 2013, identify the population of Rockcastle district in the state of Kentucky and the range of the age-adjusted mortality rate from drug poisoning.	The population of Rockcastle district in 2013 was 16744. The age-adjusted mortality rate category from drug poisoning was 24-25.9.
drugpoisoningsft_025001	Comma-structured, two-part question	What was the population of Los Alamos district in the state of New Mexico in 2000, and which range was the age-adjusted mortality rate from drug poisoning in?	The record shows a population of 18263 for Los Alamos district, New Mexico in 2000, and places the age-adjusted mortality rate from drug poisoning in the 6-7.9 category.
drugpoisoningsft_045001	District-first reversed order	What population did Covington district cover in 2011, and what was the category of the age-adjusted mortality rate from drug poisoning in that district of Mississippi state?	In 2011, the population for Covington district, Mississippi was recorded as 19509, and the age-adjusted mortality rate from drug poisoning is recorded in the 8-9.9 range.

11. Data Quality Summary

Check point	Result
Duplicate `id`	0 (all unique)
Duplicate questions (verbatim)	0 (all unique)
Duplicate answers (verbatim)	0 (all unique)
Turns per record	Always 2 (human + gpt)
Empty/null fields	`url` is always an empty string; other fields are filled
Language uniformity	100% Nepali (Devanagari)
Metadata uniformity	Most fields are constant (since this is a single-domain, single-source dataset)

12. License & Usage

Aspect	Detail
License	Public Domain – U.S. Government
License tier	Permissive (free to use)
Suitable uses	Nepali-language instruction-following / QA model fine-tuning, training single-fact retrieval capability, developing numeric-statistics comprehension ability
Limitations	Since the dataset covers only a single domain (drug poisoning mortality), it needs to be combined with additional data for general knowledge or multi-domain training

13. Summary

This dataset is built on county-level population and age-adjusted drug/substance poisoning mortality rate statistics sourced from the US CDC, and consists of 56,448 synthetic single-turn question-answer pairs prepared in Nepali. It includes multi-dimensional diversity in lexical choice, syntactic structure, and sentence mood (interrogative vs. imperative), but the subject matter is always limited to the same kind of "one state + one district + one year" population and mortality-rate-category lookup pattern (there are no comparative or multi-year questions).