|
|
--- |
|
|
language: |
|
|
- en |
|
|
- ig |
|
|
- ha |
|
|
- yo |
|
|
license: apache-2.0 |
|
|
base_model: McGill-NLP/AfriqueQwen-8B |
|
|
tags: |
|
|
- tool-calling |
|
|
- function-calling |
|
|
- igbo |
|
|
- hausa |
|
|
- yoruba |
|
|
- nigeria |
|
|
- voice-assistant |
|
|
--- |
|
|
|
|
|
# MadamLLM-Tools |
|
|
|
|
|
**Madam** is a Nigerian multilingual voice assistant fine-tuned for native tool calling across Igbo, Hausa, Yoruba, and English. |
|
|
|
|
|
This model (`madamLLM-tools`) extends [`babs/madamLLM-dpo`](https://huggingface.co/babs/madamLLM-dpo) with tool calling capabilities using Qwen3's native Hermes-style `<tool_call>` format. |
|
|
|
|
|
## Model Details |
|
|
|
|
|
- **Base model:** [McGill-NLP/AfriqueQwen-8B](https://huggingface.co/McGill-NLP/AfriqueQwen-8B) (Qwen3-8B continued pretrained on 27.5B tokens across 20 African languages) |
|
|
- **Training stages:** SFT → SimPO DPO → Tool Calling SFT (this model) |
|
|
- **Method:** LoRA SFT (rank 64, alpha 128, RSLoRA, all linear layers + lm_head) |
|
|
- **Framework:** LLaMA-Factory |
|
|
- **Hardware:** 2× NVIDIA A100 SXM 80GB |
|
|
|
|
|
## Tool Calling Evaluation |
|
|
|
|
|
| Metric | Score | |
|
|
|--------|-------| |
|
|
| Tool calling accuracy | 20/20 — **100%** | |
|
|
| Irrelevance detection | 10/10 — **100%** | |
|
|
| Args in English | 19/20 — 95% | |
|
|
| Weighted overall | **99.0%** | |
|
|
|
|
|
All four languages scored 100% on tool accuracy: English ✅ Igbo ✅ Hausa ✅ Yoruba ✅ |
|
|
|
|
|
## Supported Tools |
|
|
|
|
|
Madam knows 8 Nigeria-specific tools out of the box: |
|
|
|
|
|
| Tool | Description | |
|
|
|------|-------------| |
|
|
| `get_weather` | Current weather for Nigerian cities | |
|
|
| `get_exchange_rate` | NGN exchange rates | |
|
|
| `get_news` | Nigerian news headlines | |
|
|
| `get_market_prices` | Commodity prices (rice, garri, fuel, etc.) | |
|
|
| `search_health_info` | Basic health and first-aid guidance | |
|
|
| `get_prayer_times` | Islamic prayer times | |
|
|
| `send_message` | Send SMS/WhatsApp messages | |
|
|
| `set_reminder` | Set time-based reminders | |
|
|
|
|
|
## Tool Calling Format |
|
|
|
|
|
Uses Qwen3's native Hermes-style format. Pass tools via `apply_chat_template`: |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
import torch |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("babs/madamLLM-tools", torch_dtype=torch.bfloat16, device_map="auto") |
|
|
tokenizer = AutoTokenizer.from_pretrained("babs/madamLLM-tools") |
|
|
|
|
|
tools = [ |
|
|
{ |
|
|
"type": "function", |
|
|
"function": { |
|
|
"name": "get_weather", |
|
|
"description": "Get the current weather conditions for a city in Nigeria", |
|
|
"parameters": { |
|
|
"type": "object", |
|
|
"properties": { |
|
|
"city": {"type": "string", "description": "City name, e.g. Lagos, Kano, Abuja"} |
|
|
}, |
|
|
"required": ["city"] |
|
|
} |
|
|
} |
|
|
} |
|
|
] |
|
|
|
|
|
messages = [ |
|
|
{"role": "system", "content": "You are Madam, a Nigerian voice assistant."}, |
|
|
{"role": "user", "content": "Yaya yanayi a Kano yau?"} # Hausa: What's the weather in Kano today? |
|
|
] |
|
|
|
|
|
text = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True, enable_thinking=False) |
|
|
inputs = tokenizer(text, return_tensors="pt").to(model.device) |
|
|
output = model.generate(**inputs, max_new_tokens=200, temperature=0.3, do_sample=True, eos_token_id=[151645, 151643]) |
|
|
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)) |
|
|
# → <tool_call>{"name": "get_weather", "arguments": {"city": "Kano"}}</tool_call> |
|
|
``` |
|
|
|
|
|
## vLLM Serving |
|
|
|
|
|
```bash |
|
|
vllm serve babs/madamLLM-tools \ |
|
|
--tool-call-parser hermes \ |
|
|
--enable-auto-tool-choice \ |
|
|
--dtype bfloat16 \ |
|
|
--max-model-len 4096 |
|
|
``` |
|
|
|
|
|
## Training Data |
|
|
|
|
|
| Source | Examples | Purpose | |
|
|
|--------|----------|---------| |
|
|
| Synthetic (Groq gpt-oss-120b) | 11,924 | Nigeria-specific tool calls, 4 languages | |
|
|
| APIGen-MT-5k | 1,009 | Multi-turn tool chains | |
|
|
| xLAM-60k sample | 2,500 | Diverse tool schemas (1,240 unique tools) | |
|
|
| Hermes FC v1 | 1,100 | Format alignment | |
|
|
| SFT replay | 4,000 | Anti-forgetting (Igbo/Hausa/Yoruba/English) | |
|
|
| **Total** | **20,533** | | |
|
|
|
|
|
50% of tool-calling examples use function masking (Hammer technique) for generalization to unseen tools. |
|
|
|
|
|
## Soul |
|
|
|
|
|
Madam embodies the nine fruits of the Spirit (Galatians 5:22-23): love, joy, peace, patience, kindness, goodness, faithfulness, gentleness, and self-control. |
|
|
|
