---

license: cc-by-sa-4.0

base_model: Qwen/Qwen3-4B-Instruct-2507

base_model_relation: finetune

language:

• it

pipeline_tag: text-generation

library_name: gguf

tags:

• gguf
• llama.cpp
• travel
• rag
• lora
• pocket-travel

---

qwen3-4b-instruct-2507-travel-it-GGUF

On-device travel assistant for the Pocket Travel Android app:

a LoRA fine-tune of Qwen/Qwen3-4B-Instruct-2507, merged and quantized to GGUF Q4_K_M, that

answers in Italian, in at most 3 sentences, using only the travel-guide CONTEXT it is given, and says so

explicitly when the context does not contain the answer.

> In italiano: assistente di viaggio sul dispositivo per l'app Pocket Travel. Risponde in italiano, in massimo 3 frasi,

> solo con le informazioni della guida passate nel CONTESTO, e dice chiaramente quando mancano.

| File | Size | Quantization | Memory in use | Suggested device RAM |

|---|---|---|---|---|

| qwen3-4b-instruct-2507-travel-it-Q4_K_M.gguf | 2,382 MiB | Q4_K_M with importance matrix | about 4.6 GiB | 12 GB or more |

Memory in use: weights, the CPU copy of repacked weights, KV cache and compute buffers, measured with llama.cpp on

CPU with the app's settings (context 4096 tokens, batch 512, f16 KV cache). With about 0.4 GB for the app itself, model

and app take about 45% of the suggested device RAM or less, leaving the rest to the system and other apps.

Intended use

Retrieval-augmented answers on a phone, fully offline: the app searches the downloaded

Wikivoyage guide of a region (SQLite FTS), puts 1 to 3 sections in the prompt and asks the question. The

model is not meant as a general chatbot or knowledge source: without a context it should refuse.

The app uses this model when its interface is in Italian.

Prompt format (user turn of the model's chat template, as sent by the app):

Sei una guida turistica offline.
Rispondi in massimo 3 frasi, in italiano, usando solo le informazioni nel CONTESTO.
Se il contesto non basta, dillo esplicitamente.

CONTESTO: <guide sections>

DOMANDA: <question>

Example with llama.cpp (context 4096, as in the app):

llama-server -m qwen3-4b-instruct-2507-travel-it-Q4_K_M.gguf -c 4096

Evaluation

Held-out test set of 508 questions over 11 countries never seen in training (and their cities), phrased

outside the training templates and with contexts shaped exactly like the app's: 160 answerable questions on the

countries' guides, 44 paraphrased questions whose topic is missing from the context, 110 off-topic and 44 with

an empty context, 50 answerable and 20 unanswerable city questions, and smaller groups on the history and climate of

cities, nearby places and next departures. This model and its base model were evaluated as GGUF with llama.cpp

(greedy decoding), on the same questions and with the same prompt; the base model is the official

unsloth/Qwen3-4B-Instruct-2507-GGUF quantization (UD-Q4_K_XL) that the app offers in the same

RAM tier. Refusal rates: higher is better on the first block, lower is better on the second.

| Metric | This model | Base model (Qwen3-4B-Instruct-2507) |

|---|---|---|

| Refuses when the context is empty (no guide text found) | 100% | 100% |

| Refuses off-topic questions | 99.1% | 95.5% |

| Refuses paraphrased country questions whose topic is missing from the context | 97.7% | 70.5% |

| Refuses city questions whose topic is missing from the page | 90% | 70% |

| Refuses city history or climate questions the context does not cover | 100% | 100% |

| Refuses "what is near me" questions with no matching place nearby | 100% | 90% |

| Refuses departure questions with no matching departure | 100% | 60% |

| Wrongly refuses answerable country questions | 2.5% | 13.1% |

| Wrongly refuses answerable city questions | 2% | 10% |

| Wrongly refuses answerable city history or climate questions | 0% | 4.2% |

| Wrongly refuses answerable "what is near me" questions | 21.4% | 0% |

| Wrongly refuses answerable departure questions | 0% | 21.4% |

A refusal is any recognizable "the context does not say" answer, counted with the same rule for both models (the base

model words it in its own way). The base model's low rate of wrong refusals on some groups comes with few refusals

overall: where it does not refuse, it often answers from its own knowledge or with text from the context that does

not answer the question.

Training

• Method: LoRA (16-bit, Unsloth) merged into the base weights, 1 epoch, loss on answer tokens only, weighted by

question category.

• Data: 32,685 synthetic question/answer pairs with extractive answers, about 22% of them refusal examples, built

from the Wikimedia dumps of Oct 1, 2026: Italian Wikivoyage (country and city pages, cleaned with the same code as

the app's guides), Italian Wikipedia (country articles on cuisine, culture, telecommunications and media; history

and climate of cities), sections of English Wikivoyage machine-translated into Italian where the Italian page has

none (MarianMT Helsinki-NLP/opus-mt-tc-big-en-it, CC BY 4.0, marked as such in ATTRIBUTION.tsv), quick facts

from Wikidata (capital, currency, calling code, driving side), emergency numbers and vaccination summaries.

Contexts are shaped like the app's: 1 to 3 guide sections with distractor sections, up to 2000 characters, plus the

blocks the app adds (nearby points of interest, next departures, distances between cities, emergency numbers,

personal notes).

• Questions only (answers are always taken from the context, or are the refusal sentence): user-style travel questions

(written by a model imitating the user) from soniawmeyer/travel-conversations-finetuning

(UltraChat part, MIT), machine-translated into Italian; off-topic questions from

truthful_qa_italian (Apache 2.0) and

alpaca-cleaned-italian (CC BY 4.0 as declared by

the copy; the original Stanford Alpaca data is CC BY-NC 4.0), always with a refusal as answer.

• Quantization: llama-quantize Q4_K_M with an importance matrix calibrated on 400 training prompts.

Limitations

• Answers are only as good and as current as the guide text in the context: check prices, opening hours,

entry requirements and safety information with official sources before traveling.

• Trained to answer in Italian; questions in other languages get Italian answers.
• It sometimes answers city questions whose topic is not on the city page instead of refusing (see the table), and

it may refuse "what is near me" questions that name the place with a synonym of the label in the context.

• It may reproduce sentences of the source texts verbatim (see license below).

License and attribution

• Base model: Qwen/Qwen3-4B-Instruct-2507, Apache-2.0; its license text is included as

LICENSE-base-model.

• Training data: Wikivoyage and Wikipedia (mostly Italian, some English sections machine-translated), CC BY-SA 4.0.

Because the model can reproduce that text, the fine-tuned weights are released under CC BY-SA 4.0. The full list of source pages (title, URL, license) is in ATTRIBUTION.tsv.

• Vaccination summaries from Travel.gc.ca (OGL-Canada 2.0: contains information licensed under the Open Government

Licence - Canada) and from TravelHealthPro (UKHSA / NaTHNaC, OGL v3.0: contains public sector information licensed

under the Open Government Licence v3.0); emergency numbers from Travel.gc.ca and gov.uk (same licenses), Wikipedia,

Wikivoyage and Wikidata.

• Wikidata (quick facts, emergency numbers): CC0 1.0.
• No content from Viaggiare Sicuri (Farnesina) is included.