Koa-AI-v1-Code-3B

Koa-AI v1 Code 3B

Koa-AI-v1-Code-3B is an ultra-lightweight, high-efficiency 3B parameter model fine-tuned for code generation, multi-step agentic planning, and debugging tasks.

Built on top of mistralai/Ministral-3b-instruct using Unsloth and QLoRA, it is optimized to run blazingly fast on consumer hardware, local edge devices, and laptop GPUs without sacrificing code reasoning capabilities.

---

⚡ Highlights

• Base Model: mistralai/Ministral-3b-instruct
• Dataset: Fine-tuned on multi-turn debugging and agentic coding trace data (greghavens/fable-5-coding-and-debugging-traces).
• Efficiency: Lightweight 3B parameter size allows low-latency, real-time code completion in local IDE extensions (e.g., Continue, VS Code).
• Extended Context Support: Native context window up to 128,000 tokens (SFT fine-tuned at a 2,048 sequence length cap).

---

📊 Training Specifications

---

💻 Quickstart & Usage

Running with transformers (Python)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "vamazing/Koa-AI-v1-Code-3B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

prompt = "Write a Python function to check if a number is prime and optimize it for speed."
messages = [{"role": "user", "content": prompt}]

formatted_prompt = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True
)

inputs = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Evaluation Results

IFEval (Instruction Following Evaluation)

Note: Preliminary evaluation conducted using lm-evaluation-harness on a 100-sample test run (--limit 100).

HumanEval (Code Generation)

Note: Full zero-shot benchmark conducted using lm-evaluation-harness wrapper with an optimized vLLM fp16 inference engine on dual T4 GPU hardware.