---
license: apache-2.0
base_model: Qwen/Qwen2.5-0.5B-Instruct
base_model_relation: finetune
pipeline_tag: text-generation
library_name: llama.cpp
language: [en, cs, sk, de, fr, es, it, pt, ja, ko, zh]
tags:
  - gguf
  - qwen2.5
  - quantized
  - llama.cpp
  - llama.rn
  - on-device
  - lora
  - companion
  - text-rewriting
---

# ActiveGotchi AI — v1.0.0

A tiny multilingual **companion-voice rewriting model**. Given a short,
pre-computed activity or sleep summary (steps, minutes, hours, streaks), it
rewrites the text as a warm, playful letter from a virtual pet — in the
requested language, keeping every number, name and emoji exactly as given.

It is a narrow specialist: not a chatbot, not an assistant, not a coach, and
not a medical tool.

## Lineage

| | |
|---|---|
| Parent model | [Qwen/Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) (494M params, Apache-2.0) |
| Fine-tune | LoRA (2.93M trainable params, ~0.6%), merged into the base |
| Training | 3-stage curriculum (voice → summary rewriting → long-term reviews) + mixed rehearsal pass, on ~1M fully synthetic multilingual samples — no real user data |
| Quantization | GGUF quants of the merged fine-tune, produced with [llama.cpp](https://github.com/ggml-org/llama.cpp) `llama-quantize` |
| Released | 2026-07-19 |

## Files

| File | Size |
|---|---|
| `activegotchi-ai-v1.0.0-Q4_K_M.gguf` | 398 MB |
| `activegotchi-ai-v1.0.0-Q5_K_M.gguf` | 420 MB |

**Recommended:** `Q4_K_M` — the best size/quality balance for phones and
other memory-constrained devices.

## Usage

ChatML prompt format, built into the GGUF chat template — llama.cpp,
llama.rn, LM Studio, Ollama etc. apply it automatically. Send ONE user
message that states the persona/task and ends with the text to rewrite:

```
You are a small cheerful companion. Rewrite the following daily summary for
your human in your own voice. Keep every number exactly as given, keep it to
3-4 short sentences. Write your entire reply in Czech. Do not use any other
language. SUMMARY TO REWRITE: <template letter with the real numbers>
```

Suggested inference settings: `temperature 0.6`, `max tokens 220`,
context 2048.

```sh
llama-cli -m activegotchi-ai-v1.0.0-Q4_K_M.gguf -st \
  -p "<prompt as above>" -n 220 --temp 0.6
```

## Trained behavior

- Never emits a number that is not present in the prompt.
- Replies only in the language the prompt pins.
- Short letters (3–6 sentences), no greetings, no filler.
- Warm, non-judgmental tone; no medical advice, diagnosis or shaming.

## Limitations

The model only knows what the prompt contains — it has no memory, no health
knowledge, and no general-assistant abilities. Outside its rewrite task,
output quality is undefined. Languages beyond en/cs/de/es/fr received less
training weight. Synthetic-data style ceiling applies.

## Release notes

Automated release from run_full_flow_mac.sh