Sign in to view source links and access this dataset
Description
10,000 Hindi instruction–response pairs used to train the Qwen3-4B-Hindi-Instruct-v2 model. The dataset is a reproducible recipe derived from a larger 50,000-row source dataset created by FreedomIntelligence as part of the MultilingualSIFT project. It was last updated on May 30, 2026.
Use Cases
Instruction tuning of Hindi language models based on the described instruction–response pairs.
Benchmarking model performance on Hindi instruction-following tasks.
Research on cross-lingual transfer learning using a curated Hindi subset.
Fine-tuning large language models for specific Hindi-language applications.
Strengths
Contains exactly 10,000 instruction–response pairs, providing a defined scale.
Derived from a larger, credited source dataset of approximately 50,000 rows.
Documented as a reproducible recipe for training a specific model (Qwen3-4B-Hindi-Instruct-v2).
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count for the final dataset is known, but the exact filtering criteria are not fully detailed here.
The description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
FreedomIntelligence/alpaca-gpt4-hindi
Collection Method
Filtered and curated from a larger source dataset.
Freshness
Last updated 2026-05-30 17:44:50; freshness should be verified.
License is unknown; terms of use must be verified before application.