File size: 2,914 Bytes
a99c714 18ff929 a99c714 18ff929 ba88c00 a99c714 ba88c00 a99c714 18ff929 ba88c00 18ff929 a99c714 ba88c00 a99c714 ba88c00 a99c714 ba88c00 a99c714 ba88c00 a99c714 ba88c00 a99c714 ba88c00 a99c714 ba88c00 a99c714 ba88c00 a99c714 ba88c00 a99c714 ba88c00 a99c714 6bf588f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 | ---
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
|