Circuit

Fine-tuned Phi-3 for Logical Reasoning

Model performance

Benchmark

Trained on the lucasmccabe/logiqa dataset, Circuit enhances the model’s ability to reason through complex problems, answer multi-step logic questions, and provide consistent explanations.

Model Details

Property	Value
Base model	microsoft/Phi-3-mini-4k-instruct
Fine-tuned for	Logical Reasoning
Dataset	lucasmccabe/logiqa
Technique	LoRA fine-tuning, merged for direct use
Formats available	Full (HF Transformers) + Quantized (.gguf for llama.cpp / Ollama)
Project	Circuit
Fine-tuned by	Rudransh

Model Variants

Variant	Description	File
Full model	Merged LoRA with base, compatible with transformers	pytorch_model.bin
Quantized model (GGUF)	Optimized for CPU/GPU inference via llama.cpp, text-generation-webui, or Ollama	circuit_phi3_q4.gguf

Example Usage (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "rudranshjoshi/circuit",
    torch_dtype=torch.float16,
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "rudranshjoshi/circuit",
    trust_remote_code=True
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

prompt = "Your prompt here"
inputs = tokenizer(prompt, return_tensors="pt").to(device)
outputs = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Summary

Base model: Phi-3 Mini 4K Instruct

Dataset: LogiQA (lucasmccabe/logiqa)

Training method: LoRA fine-tuning, later merged

Hardware: NVIDIA RTX 1080

Epochs: ~3

Objective: Improve reasoning consistency and structured explanations

Acknowledgements

Microsoft for Phi-3

Lucas McCabe for LogiQA dataset

Fine-tuned and quantized by Rudransh under Project Circuit