---
base_model: Qwen/Qwen2.5-1.5B-Instruct
license: apache-2.0
language:
- fr
- en
tags:
- gguf
- ollama
- qwen2.5
- navigation
library_name: gguf
---

# teleport-nav — Q8_0 GGUF (Qwen2.5-1.5B)

Fine-tuned **Qwen2.5-1.5B-Instruct** that turns French/English natural-language
navigation requests into the strict scenario JSON consumed by the **Teleport
Control** app, and refuses off-domain requests as valid empty-step JSON.

| | |
|---|---|
| Base model | `Qwen/Qwen2.5-1.5B-Instruct` (Apache-2.0) |
| Quantization | `Q8_0` (~1.6 GB) |
| Languages | French + English |
| Prompt format | Qwen2.5 ChatML |
| Sampling | `temperature 0.3`, `format: json` |
| Release eval | 150 held-out prompts, 100% valid JSON / 100% step type / 100% guardrail |
| GGUF SHA256 | `a435608db1988fbc8b262276c1e1b359ffbfc4b1433d7db30a27b8736a061b13` |

Trained for **multi-destination chains** ("va à X, puis Y, ensuite Z, puis
rentre"), timed pauses (`stayAt`), returns (`returnToOrigin`), transport modes,
and robustness to messy real-world phrasing (lowercase, no accents, abbreviations).
Step `type` is restricted to `moveTo, moveToCategory, moveDirection, stayAt,
returnToOrigin`.

**System prompt (train == serve):**

> `Tu es l'expert de navigation de Teleport Control. Tu traduis les requêtes en scénarios JSON stricts.`

## Run with Ollama

```bash
ollama pull hf.co/lxucan/teleport-nav:Q8_0
ollama run hf.co/lxucan/teleport-nav:Q8_0 "va à la boulangerie, puis à la pharmacie, et rentre"
```

Output is strict JSON: `{name, description, isLooping, steps:[{type, mode, label, …}]}`.

## Validation

The release GGUF was validated through Ollama as `teleport-nav-1.5b` on the
held-out `finetune/data/test.jsonl` set:

| Metric | Score |
|---|---:|
| JSON validity | 100% |
| Step type correctness | 100% |
| Guardrail recall | 100% |
| Coordinates / direction / road style / transport buckets | 100% |
| Target correctness | 89.33% |
| Overall strict scorer | 78.67% |