phi-4-mini-instruct-4b-usm-tau-py-0003/README.md

tags, datasets, language, base_model, pipeline_tag, license

tags	datasets	language	base_model	pipeline_tag	license
gguf llama.cpp unsloth ollama phi4 phi-4 phi-4-mini phi4mini phi-4mini 4b python code coder python-code tlet-ai tlet	iamtarun/python_code_instructions_18k_alpaca	en	unsloth/Phi-4-mini-instruct-unsloth-bnb-4bit	text-generation	mit

phi-4-mini-instruct-4b-usm-tau-py-0003 (GGUF)

Internal Model Name: Tau0-Py-003-4B-ir

Explaining how TLET AI public & internal model names work:

• Public Model Name: {basemodel}-{parameters}-{type}-{series}-{hyperspecialization (when USM type)}-{release type number}{versioning number (digits are 3x what release type number digits are)}
• So: phi-4-mini-instruct-4b-usm-tau-py-0003 means:
• • Base: phi-4-mini-instruct (specifically unsloth/Phi-4-mini-instruct-unsloth-bnb-4bit)
  • Parameters: 4 Billion
  • Type: USM (Ultra Specialized Model)
  • Series: Tau (Proof of Concept model series)
  • Hyperspecialization (is USM): Python
  • Release Type: 0, Proof of Concept/Early Works stage
  • Release Version: 003, this is the 3rd model in the Tau series of models (includes fine-tunes that come from other base models.)
• Internal/Private Model Name: {series}{release type number}-{hyperspecialization}-{versioning number (digits are 3x what release type number digits are)}-{parameters}-{ir IF model is Inference Ready (Ollama)}
• So: Tau0-Py-003-4B-ir means:
• • Series: Tau (Proof of Concept model series)
  • Release Type: 0, Proof of Concept/Early Works stage
  • Hyperspecialization (is USM): Python
  • Release Version: 003, this is the 3rd model in the Tau series of models (includes fine-tunes that come from other base models.)
  • Is inference ready.

Ollama Commands (with recommended Q5_K_M quantization)

Pull

ollama pull hf.co/tletai/phi-4-mini-instruct-4b-usm-tau-py-0003:Q5_K_M

Run Command

ollama run hf.co/tletai/phi-4-mini-instruct-4b-usm-tau-py-0003:Q5_K_M

Aliasing

While you are running it, you can run the following command to save it with it's much simpler internal name for ease of use via Ollama. (This command should be ran AFTER running ollama run, meaning, when you are already chatting with the model in the CLI.)

>>> /save tau0-py-003-4B-ir:Q5_K_M

This will allow you to run it using this command instead:

ollama run tau0-py-003-4B-ir:Q5_K_M

You can remove parts you don't want, such as "-4B-ir" or ":Q5_K_M" (which isn't really needed if you're just planning on downloading a single quantization anyways) from the /save command as you wish.

Deleting Aliases

This command will remove the alias but keep the model:

ollama rm tau0-py-003-4B-ir:Q5_K_M

Complete Deletion

To fully remove the model, remove all aliases and also remove the original pull:

ollama rm hf.co/tletai/phi-4-mini-instruct-4b-usm-tau-py-0003:Q5_K_M

Fine-tuning

• Base model: unsloth/Phi-4-mini-instruct-unsloth-bnb-4bit
• Done using QLoRA, Paged AdamW 32-bit with an 8K Context Length on an NVidia RTX 3060 Ti (8GB VRAM) for 6h47m at about ~112W power draw for most of the time, with occassional hike-ups to at-most 217W (of therotical possible 225W.)
• Tokens during training: 4699303.
• Epochs completed: 1.33 (67% of 2 target epochs, runtime got too long for proof of concept, so it was cancelled. Steps for more precise, in our config: 3099/4654.)
• Done using Unsloth Studio, which largely increased training efficency and speed.
• If you need specifics for research purposes, possible collaboration, fine-tuning a model yourself or are just curious, feel free to reach out. We do not have specific, timed power usage data anymore. It was discarded immediately after it was used, do not ask for it.

System Specifications

• CPU 1x i9-12900KF
• RAM 4x 16GB of RAM (DDR4, 3600MHz)
• • TOTAL 64GB of RAM (DDR4, 3600MHz)
• GPUS
• • 1x NVIDIA GeForce RTX 3060 Ti (8GB of VRAM)
• OS Windows 10 Native