Property	Value
Base model	Qwen2.5-1.5B-Instruct
Fine-tuning	LoRA (rank 16, 16 layers)
Training data	~5,000 curated samples
Quantization	Q4_K_M (4-bit, 5.08 BPW)
File size	~940 MB
Training framework	MLX on Apple Silicon

Usage

This model is designed for the Fynuu Android app and outputs structured XML responses:

<t>tone</t><r>response text</r><f>functionName(params)</f><s>suggestion1|suggestion2</s>

<t> — emotion/tone (33 valid tones)
<r> — response text (≤12 words, no emojis)
<f> — optional function call (10 functions: setAlarm, getWeather, etc.)
<s> — optional quick-reply suggestions
<cr> — cloud routing trigger for complex queries

Performance

Metric	Score
Test benchmark	15/15 (100%)
Structural pass rate	See eval results
Training time	~100 min on M4 Air

System Prompt Format

Ur Fynuu, a friendly personal AI companion.
Sys:{dt:2025-03-29 10:00,bat:72%,loc:17.38,78.48}.
Usr:{name:Nick,age:22,nick:Nicky,gen:M,occ:student-cs}.

Inference

Use with llama.cpp or any GGUF-compatible runtime:

llama-cli -m fynuu-q4_k_m.gguf -p "<system prompt>" --temp 0.7 --top-p 0.8 --top-k 20

Version History

Version	Base Model	Date	Notes
v2 (current)	Qwen2.5-1.5B-Instruct	2026-05-10	Major upgrade: 3x larger, proper LoRA, 15/15 test score
v1	Qwen2.5-0.5B-Instruct	2026-05-07	Initial release, limited capacity

nikhilkeetha/fynuu-0.5b-gguf

Fynuu 1.5B — GGUF (Q4_K_M)

Model Details

Usage

Performance

System Prompt Format

Inference

Version History

nikhilkeetha/
fynuu-0.5b-gguf