Somtharu181coder/educational_vedantu_FAQs_Nepali_sft_dataset

Vedantu CUET FAQs — Nepali SFT Dataset

A single-turn instruction-following (Q&A) dataset in Nepali, built from Vedantu's CUET (Common University Entrance Test) 2026 FAQ page. The dataset is formatted for supervised fine-tuning (SFT) in a conversations-style chat schema.

Summary

Records	49
Language	Nepali (ne / npi)
Script	Devanagari (Deva)
Format	JSONL, conversations (human/human turns)
Domain	Education — CUET exam FAQs
Task type	Instruction-following (single-turn Q&A)
Source	Vedantu (vedantu.com/cuet/cuet-faqs)

How this dataset was built

1. Merge — Two raw scrapes of the same Vedantu FAQ page were combined:

2. Semantic review & cleaning

   • Removed 75 rejected entries that were exact duplicates of content already captured elsewhere (repeat scraping passes).
   • Deduplicated the remaining pool by normalized question text.
   • Manually inspected and dropped entries where the question/answer boundary had clearly shifted during scraping (e.g. a question about "NTA Score Validity" paired with an unrelated answer about "DigiLocker") — a scraping artifact, not a real FAQ.
   • Repaired run-together text caused by missing whitespace in scraped tables/lists (e.g. PwD categories, exam-pattern tables, admit-card detail lists).
   • Result: 49 clean, unique, correctly-paired Q&A items.

3. Translation — Each Q&A pair was translated from English into natural, formal Nepali (Devanagari script), preserving exam-specific terms (NTA, CUET, PwD, DigiLocker, etc.) where an English acronym is the standard term used by Nepali speakers as well.

4. Formatting — Records were converted into the target SFT schema (see below). The schema's structure was reused as specified; field values were populated to accurately describe this dataset rather than copied from any reference example.

Schema

Each record is a JSONL object:

  "id": "sg_<32-char hash>",
  "conversations": [
    { "from": "human", "value": "<question in Nepali>" },
    { "from": "gpt",   "value": "<answer in Nepali>" }
  ],
  "source": "vedantu/cuet_faqs:merged:train",
  "source_name": "vedantu_cuet_faqs_nepali",
  "source_repo": "vedantu/cuet_faqs",
  "source_config": "merged",
  "source_split": "train",
  "source_revision": "local-merge-2026-09-07",
  "source_row_id": "vedantu:cuet_faqs:merged:<n>",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "Unspecified",
  "license_tier": "unspecified",
  "task_type": "instruction-following",
  "generation_type": "machine-translated",
  "condition": "web-scraped,machine-translated",
  "url": "https://www.vedantu.com/cuet/cuet-faqs",
  "metadata_json": "{ ...see below... }"

Field notes

• id — Stable synthetic ID (sg_ + MD5 hash of the record index and original English question). Unique per row.
• conversations — Exactly one human turn (the question) and one human turn (the answer). No multi-turn dialogue.
• source / source_name / source_repo / source_config / source_split — Provenance identifiers for this merged/translated build; not a published HF dataset.
• source_revision — Local build tag, since this isn't pinned to a versioned upstream repo.
• source_row_id — Sequential row identifier within this merge (vedantu:cuet_faqs:merged:<n>).
• language / language_code / script — Nepali, ISO 639-3 npi, Devanagari.
• license / license_tier — license: cc-by-4.0 and license_tier: permissive
• task_type — instruction-following, consistent with single-turn FAQ-style Q&A.
• generation_type — machine-translated: the underlying FAQ content is human-authored (scraped), but the Nepali text was produced by machine translation, not by a human Nepali speaker.
• condition — web-scraped,machine-translated, describing the full provenance chain.
• url — Original source page for all records.
• metadata_json — A JSON-encoded string containing:
  • domain: "education"
  • category: "CUET"
  • original_language: "English"
  • original_question / original_answer: the source English text, kept for traceability/back-translation checks
  • merge_note: brief description of the merge/cleaning process

Known limitations

• Machine translation, not human-reviewed Nepali. Spot-check before using for anything user-facing; exam terminology (fee amounts, dates, section names) should be verified against an authoritative CUET source before relying on it for real guidance.
• Time-sensitive content. Several answers contain exam-cycle-specific details (fees, dates, correction windows) tied to the "2026" cycle referenced on the source page. These will go stale and should not be treated as current CUET information.
• Not de-duplicated against other datasets. Only internal duplication (between the accepted/rejected scrape files) was removed.
• Single-turn only. No multi-turn conversations, follow-ups, or clarification exchanges.
• Answer length varies widely (~40–1,300 characters) — a few answers (e.g. the CUET 2026 exam-pattern breakdown) are long, table-derived explanations rather than short FAQ answers.

Suggested uses

• Fine-tuning or evaluating Nepali-language instruction-following models on short factual Q&A.
• Domain-adaptation for education/exam-assistant use cases in Nepali.
• As a small seed set to bootstrap a larger Nepali CUET/exam-FAQ corpus (e.g. combined with other exam boards or translated into additional South Asian languages).

File

• vedantu_cuet_faqs_nepali_sft.json — 49 records, UTF-8 JSON array, ~1,470 lines pretty-printed.