alok21in/llama-3.1-8b-function-calling

🦙 Llama 3.1 8B — Function Calling (LoRA)

Fine-tuned by Alok Kumar Dubey | GitHub | LinkedIn | HuggingFace

A LoRA adapter fine-tuned on top of Meta-Llama-3.1-8B-Instruct for function calling and tool use. The model learns to identify when to call a function, generate correct JSON arguments, and formulate a natural language response from the function output.

---

🚀 Quick Start

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name     = "alok21in/llama-3.1-8b-function-calling",
    max_seq_length = 1024,
    load_in_4bit   = True,
)
FastLanguageModel.for_inference(model)

prompt = """Below is a system context describing available tools, followed by a user request.

### Available Tools:
You have access to a weather API that takes a city name and returns current weather.

### Conversation & Response:
USER: What is the weather in London?
ASSISTANT:"""

inputs  = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200, pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

📊 Example Output

Input:

USER: What is the weather in London?

Output:

<functioncall> {"name": "get_weather", "arguments": '{"city": "London"}'}

FUNCTION RESPONSE: {"city": "London", "weather": "Sunny", "temperature": 22, "humidity": 60}

The weather in London is sunny, with a temperature of 22°C and humidity of 60%.

---

🏋️ Training Details

Parameter	Value
Base Model	meta-llama/Meta-Llama-3.1-8B-Instruct
Dataset	glaiveai/glaive-function-calling-v2
Training Samples	5,000
LoRA Rank (r)	8
LoRA Alpha	16
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Quantization	4-bit (QLoRA)
Max Seq Length	1024
Batch Size	1 (effective 8 with grad accumulation)
Learning Rate	2e-4
Steps	60
Optimizer	adamw_8bit
Hardware	Kaggle T4 GPU (15GB)
Framework	Unsloth 2026.4.2 + TRL

---

📉 Training Loss

Step	Loss
1	1.1046
10	0.6266
30	0.4327
60	0.6127

Loss decreased from 1.10 → ~0.4 showing successful learning.

---

⚠️ Limitations

• Trained on only 5,000 samples — a larger run would improve reliability
• Only 60 training steps — more steps would reduce loss further
• Best used as a research/portfolio demonstration

---

📜 License

This model is built on Meta's Llama 3.1 and follows the Llama 3 Community License.

---

🙏 Acknowledgements

• Unsloth for making fine-tuning accessible on consumer GPUs
• Glaive AI for the function calling dataset
• Kaggle for free GPU compute