akadriu/albanian-news-instructions

Albanian News Instructions

The exact instruction/response pairs used to fine-tune akadriu/albanian-news-llm.

Constructed from 50,000 unique Albanian news articles sampled from akadriu/lajme-shqip by applying 5 instruction templates.

---

Splits

Split	Examples
train	50,000
validation	2,000

---

Schema

Each row is a chat conversation in standard messages format:

{
  "messages": [
    {"role": "system",    "content": "Ju jeni një asistent inteligjent i lajmeve shqiptare. ..."},
    {"role": "user",      "content": "Çfarë ndodhi sipas lajmit: '<TITLE>'?"},
    {"role": "assistant", "content": "<SUMMARY of the article in Albanian news style>"}
  ]
}

Compatible with most SFT trainers (TRL, Unsloth, Axolotl, Tinker, etc.) that accept the chat-format conversational schema.

---

Construction

Built by running prepare_training_data.py against the cleaned articles in akadriu/lajme-shqip. Steps:

1. Fetch unique articles (DISTINCT on title, summary length > 50 chars)
2. Sample 52,000 articles deterministically (seed=42)
3. For each article, pick one of 5 instruction templates randomly
4. Build a system+user+assistant chat triple
5. Shuffle, split 50,000 train / 2,000 validation

The 5 instruction templates

"Jep një përmbledhje të shkurtër të këtij lajmi: {title}"
"Çfarë ndodhi sipas lajmit: '{title}'?"
"Ky lajm i kategorisë '{category}' ka titullin: '{title}'. Çfarë thuhet në të?"
"Çfarë raportoi portali {portal} më {date} për: '{title}'?"
"{title}"   # bare title → expand to news summary

The assistant response in every example is the cleaned summary from the source article (no fabrication — direct mapping from real news data).

System prompt (same for all examples)

"Ju jeni një asistent inteligjent i lajmeve shqiptare.
 Përgjigjuni pyetjeve bazuar në lajmet dhe informacionin e disponueshëm."

---

Intended use

• Supervised fine-tuning of base LLMs to produce Albanian news prose
• Benchmarking other Albanian LLMs on news-style summarization
• Research on instruction tuning for low-resource languages
• Reproducing akadriu/albanian-news-llm

---

Honest limitations

• Only 5 templates — somewhat repetitive instruction phrasings. A more diverse template set would yield a more general instruction-following model.
• Title → summary mapping — the assistant response is always a real article summary. There are no explicit "I don't know" examples, multi-turn dialogs, or strict-format outputs.
• Single-paragraph training — no long-essay structure.
• Derivative — anyone with the source dataset can regenerate this exactly.
• News-domain bias — politics, current events, sports, entertainment. Not balanced for other domains.

For factually grounded Q&A behavior, this dataset alone is not enough — combine the resulting model with retrieval (RAG) or augment with explicit Q&A pairs and refusal examples.

---

Source data

Source dataset	akadriu/lajme-shqip
Source size	~1.19M Albanian news articles
Source portals	100+ across Albania, Kosovo, North Macedonia, and the diaspora
Cleaning applied	WordPress footers, HTML, truncation markers stripped

---

Citation

@misc{albanian_news_instructions_2026,
  title        = {Albanian News Instructions: Instruction-tuning pairs for Albanian news LLM},
  author       = {Kadriu, Arbana},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/akadriu/albanian-news-instructions}
}

---

Related resources

• 🤖 Model trained from this dataset: akadriu/albanian-news-llm
• 📰 Source articles: akadriu/lajme-shqip
• 📚 Albanian dictionary: akadriu/fjalori-shqip