Keep K3 suite selection and report-schema scoring in bash, merge K3/vision dataset_args into dpv4 yamls, and pin EvalScope at 735d920ee911 with local patches. Co-authored-by: Cursor <cursoragent@cursor.com>
3.7 KiB
3.7 KiB
Sanskriti
Overview
Sanskriti is a multiple-choice trivia benchmark testing knowledge of Indian states' culture, history,
and geography, sourced from state-specific attributes (art, cuisine, festivals, etc.) with
Wikipedia-backed answers. From the SANSKRITI paper (arXiv:2506.15355); this adapter loads the
dataset mirrored to ModelScope as evalscope/Sanskriti.
Task Description
- Task Type: Multiple-Choice Trivia Question Answering
- Input: A question about a specific Indian state's culture/geography/history, with 4 answer choices
- Output: Correct answer letter
- Subsets:
association(state-attribute association trivia),country(country-level trivia),gk(general knowledge),states(state-identification trivia)
Evaluation Notes
- Default configuration uses 0-shot evaluation (the dataset's only split, named
trainupstream despite being evaluation data) - Questions and choices are in English
- The paper acknowledges some questions involve ambiguous cultural elements; a small number of rows
(~0.6%) whose
answerdoesn't match any of the 4 listed options are skipped at load time
Properties
| Property | Value |
|---|---|
| Benchmark Name | sanskriti |
| Dataset ID | evalscope/Sanskriti |
| Paper | N/A |
| Tags | Knowledge, MCQ |
| Metrics | accuracy |
| Default Shots | 0-shot |
| Evaluation Split | train |
Data Statistics
| Metric | Value |
|---|---|
| Total Samples | 21,726 |
| Prompt Length (Mean) | 322.93 chars |
| Prompt Length (Min/Max) | 256 / 636 chars |
Per-Subset Statistics:
| Subset | Samples | Prompt Mean | Prompt Min | Prompt Max |
|---|---|---|---|---|
association |
5,453 | 343.41 | 273 | 523 |
country |
5,563 | 284.48 | 256 | 417 |
gk |
5,328 | 346.94 | 263 | 547 |
states |
5,382 | 318.17 | 260 | 636 |
Sample Example
Subset: association
{
"input": [
{
"id": "0629b222",
"content": "Answer the following multiple choice question. The entire content of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of A,B,C,D.\n\nWhich of the given regions is home to the Jarawa body painting?\n\nA) Surguja district\nB) South Andaman and Middle Andaman Islands\nC) Buddha Marg, Patna\nD) Telangana"
}
],
"choices": [
"Surguja district",
"South Andaman and Middle Andaman Islands",
"Buddha Marg, Patna",
"Telangana"
],
"target": "B",
"id": 0,
"group_id": 0,
"subset_key": "association",
"metadata": {
"state": "Andaman_and_Nicobar",
"attribute": "Art"
}
}
Prompt Template
Prompt Template:
Answer the following multiple choice question. The entire content of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of {letters}.
{question}
{choices}
Usage
Using CLI
evalscope eval \
--model YOUR_MODEL \
--api-url OPENAI_API_COMPAT_URL \
--api-key EMPTY_TOKEN \
--datasets sanskriti \
--limit 10 # Remove this line for formal evaluation
Using Python
from evalscope import run_task
from evalscope.config import TaskConfig
task_cfg = TaskConfig(
model='YOUR_MODEL',
api_url='OPENAI_API_COMPAT_URL',
api_key='EMPTY_TOKEN',
datasets=['sanskriti'],
dataset_args={
'sanskriti': {
# subset_list: ['association', 'country', 'gk'] # optional, evaluate specific subsets
}
},
limit=10, # Remove this line for formal evaluation
)
run_task(task_cfg=task_cfg)