A compact, edge-deployable group chat routing model fine-tuned on Qwen3-0.6B. Given a 5-message history window and up to 3 unprocessed candidate messages, it outputs a single structured routing decision in under 4 seconds on CPU — no response generation, pure intent classification.
Part of the Auroic project — a personal AI assistant for Indian group and personal chats.
What it does
The router sits at the front of a conversational AI pipeline. It reads the recent chat context and decides what action to take next — so downstream models only activate when actually needed.
H1: bhai sun
H2: kya hua bata
H3: okay
H4: arre
H5: haan bol
C1: ...
C2: ...
C3: bhai parents divorce ho raha hai adjust karna mushkil hai kya karoon
→ R: TYPE=text | TARGET=C3 | EFFORT=high
Always provide exactly 5 history messages and 3 candidates. Use ... for empty slots — never omit them.
Quickstart — Ollama
bash
1ollama create auroic-router-0.6b -f Modelfile
2ollama run auroic-router-0.6b
Important — the router is stateless. Every call must be a fresh context with no conversation history. In the Ollama CLI use /clear between calls. In production call the API directly:
python
1response = client.chat.completions.create(2 model="auroic-router-0.6b",3 messages=[4{"role":"system","content":"You are the Auroic Router. Given history messages H1-H5 and candidate messages C1-C3, output exactly one routing decision."},5{"role":"user","content": formatted_window}# fresh window every call6],7)
INPUT:
H1: sahi hai H2: haan bhai H3: okay H4: chal H5: theek hai
C1: ...
C2: bhai salary negotiate kaise karoon first time offer mein help
C3: ...
OUTPUT: R: TYPE=text | TARGET=C2 | EFFORT=medium
INPUT:
H1: bhai hungry hun H2: same yaar H3: kuch khate hain H4: haan H5: chal
C1: ... C2: ...
C3: bhai biryani ki yaad aa rahi hai mummy wali ghar ki bahut miss kar raha
OUTPUT: R: TYPE=media | TARGET=C3 | TITLE=biryani craving
INPUT:
H1: haha H2: bhai sach mein H3: lol H4: no way H5: 💀
C1: ...
C2: teacher ne galti se apna tiktok projector pe chala diya class mein
C3: ...
OUTPUT: R: TYPE=media | TARGET=C2 | TITLE=tiktok projector fail
INPUT:
H1: bhai sun H2: haan H3: bata H4: okay H5: hmm
C1: ...
C2: @BOT yaar dost ne 3 baje uthke help kiya exam ke liye true friendship hai
C3: ...
OUTPUT: R: TYPE=react | TARGET=C2 | TITLE=❤
Thinking Mode
The model uses Qwen3's native thinking capability. For ambiguous inputs it reasons through the context before deciding:
<think>
H1-H5 are casual affirmations. C2 asks 'salary negotiate kaise karoon first time' —
clear, actionable life advice needed. Text with medium effort because it requires
step-by-step guidance, not just emojis or media.
</think>
R: TYPE=text | TARGET=C2 | EFFORT=medium
For obvious cases (clear ignore, clear media) it skips thinking entirely and outputs the decision directly — keeping latency low where reasoning isn't needed.
Thinking enabled (default): better accuracy on ambiguous cases, ~3-4s on CPU
Thinking disabled: faster, minor accuracy drop on edge cases
To disable thinking add this to your Modelfile template:
<|im_start|>assistant
<think>
</think>
Recommended Inference Settings
temperature 0.3 (deterministic routing)
top_p 0.95
top_k 20
repeat_penalty 1.1
Model Details
Property
Value
Base model
unsloth/Qwen3-0.6B
Total parameters
616M
Trainable parameters
20.2M (3.28%)
Fine-tuning framework
Unsloth + TRL SFTTrainer
LoRA rank
r=32, alpha=32
Target modules
q/k/v/o/gate/up/down proj
Epochs
2
Effective batch size
16 (batch=2, grad_accum=8)
Learning rate
2e-4 cosine
Max sequence length
2048
Training samples
9,300
Final training loss
0.667
Quantization
Q8_0 GGUF
Hardware trained on
NVIDIA T4 16GB (Google Colab)
Training time
~44 minutes
Dataset — v4
9,300 samples of Indian group chat routing scenarios:
Split
Count
Normal windows
7,000
@BOT mention windows
1,500
Filler/sparse windows
800
Type
Count
%
text
4,042
43.5%
ignore
1,942
20.9%
react
1,688
18.2%
media
1,628
17.5%
Language distribution: 58.8% Hinglish, 23% English, 18.2% mixed
Think blocks annotated by qwen/qwen3-next-80b-a3b-instruct on NVIDIA NIM — hard tier (39.3%) gets full reasoning, medium tier (32.8%) gets short reasoning, easy tier (27.8%) gets no think block.
