AKSHIT_05/ phi15:latest

65 5 months ago

A high-quality 8-bit (Q8_0) quantization of Microsoft's Phi-1.5 (1.3B). Optimized for local coding assistance and logical reasoning with a tiny 1.8GB memory footprint. Features custom stop tokens to prevent base-model hallucinations. 

ollama run AKSHIT_05/phi15

Details

5 months ago

5125c4442a4f · 1.5GB

model
phi2
·
1.42B
·
Q8_0
template
User: {{ .Prompt }} Assistant:
system
You are a helpful, concise chat assistant. Do not generate dialogue for the User.
params
{ "stop": [ "User:", "Assistant:", "Student:" ], "temperature":

Readme

Phi-1.5 8-bit (GGUF)

This is a quantized version of Microsoft’s Phi-1.5, a 1.3 billion parameter transformer model. It was trained on “textbook-quality” data, making it exceptionally good at common-sense reasoning and Python coding despite its small size.

Features

  • Quantization: Q8_0 (8-bit) for near-original precision.
  • Size: ~1.6GB (Fits in almost any GPU or 4GB+ RAM system).
  • Optimization: Custom configuration to stop the “Base Model” from talking to itself or generating endless transcripts.

How to use

This model is configured to work as a direct chat assistant. It responds best to concise instructions.

Example Prompt:

User: Write a Python function to check if a number is prime. Assistant: [Code output]

Technical Specs

  • Context Window: 2048 tokens
  • Temperature Recommendation: 0.2 - 0.5 (Lower is better for logic/code)
  • Architecture: Refined Transformer

Modelfile Configuration

If you are building this manually, the following parameters were used to ensure stability: - PARAMETER temperature 0.2 - PARAMETER stop User: - PARAMETER stop Assistant: - TEMPLATE "User: {{ .Prompt }} Assistant: "

License

This model is released under the MIT License by Microsoft.