README.md

---
language:
- en
license: apache-2.0
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- gguf
base_model: unsloth/phi-4-unsloth-bnb-4bit
datasets:
- bespokelabs/Bespoke-Stratos-17k
- bespokelabs/Bespoke-Stratos-35k
- NovaSky-AI/Sky-T1_data_17k
- Quazim0t0/BenfordsLawReasoningJSON
- open-thoughts/OpenThoughts-114k
model-index:
- name: Phi4.Turn.R1Distill_v1.5.1-Tensors
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: IFEval (0-Shot)
      type: HuggingFaceH4/ifeval
      args:
        num_few_shot: 0
    metrics:
    - type: inst_level_strict_acc and prompt_level_strict_acc
      value: 29.95
      name: strict accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Quazim0t0/Phi4.Turn.R1Distill_v1.5.1-Tensors
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: BBH (3-Shot)
      type: BBH
      args:
        num_few_shot: 3
    metrics:
    - type: acc_norm
      value: 49.22
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Quazim0t0/Phi4.Turn.R1Distill_v1.5.1-Tensors
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MATH Lvl 5 (4-Shot)
      type: hendrycks/competition_math
      args:
        num_few_shot: 4
    metrics:
    - type: exact_match
      value: 1.59
      name: exact match
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Quazim0t0/Phi4.Turn.R1Distill_v1.5.1-Tensors
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: GPQA (0-shot)
      type: Idavidrein/gpqa
      args:
        num_few_shot: 0
    metrics:
    - type: acc_norm
      value: 2.46
      name: acc_norm
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Quazim0t0/Phi4.Turn.R1Distill_v1.5.1-Tensors
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MuSR (0-shot)
      type: TAUR-Lab/MuSR
      args:
        num_few_shot: 0
    metrics:
    - type: acc_norm
      value: 7.04
      name: acc_norm
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Quazim0t0/Phi4.Turn.R1Distill_v1.5.1-Tensors
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU-PRO (5-shot)
      type: TIGER-Lab/MMLU-Pro
      config: main
      split: test
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 45.75
      name: accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Quazim0t0/Phi4.Turn.R1Distill_v1.5.1-Tensors
      name: Open LLM Leaderboard
---

# TurnPhi Project

- **Developed by:** Quazim0t0
- **Finetuned from model :** unsloth/phi-4-unsloth-bnb-4bit
- **GGUF**
- **Trained for 8 Hours on A800 with the Bespoke Stratos 17k Dataset.**
- **Trained for 6 Hours on A800 with the Bespoke Stratos 35k Dataset.**
- **Trained for 2 Hours on A800 with the Benford's Law Reasoning Small 430 Row Dataset, ensuring no overfitting.**
- **Trained for 4 Hours on A800 with the Sky-T1_data_17k Dataset**
- **Trained for 6 Hours on A800 with the Openthoughts 114k Dataset.**
- **18$ Training...I'm actually amazed by the results.**

# OpenWeb UI Function
If using this model for Open WebUI here is a simple function to organize the models responses: https://openwebui.com/f/quaz93/phi4_turn_r1_distill_thought_function_v1

# Phi4 Turn R1Distill LoRA Adapters

## Overview
These **LoRA adapters** were trained using diverse **reasoning datasets** that incorporate structured **Thought** and **Solution** responses to enhance logical inference. This project was designed to **test the R1 dataset** on **Phi-4**, aiming to create a **lightweight, fast, and efficient reasoning model**.

All adapters were fine-tuned using an **NVIDIA A800 GPU**, ensuring high performance and compatibility for continued training, merging, or direct deployment.  
As part of an open-source initiative, all resources are made **publicly available** for unrestricted research and development.

---

## LoRA Adapters  
Below are the currently available LoRA fine-tuned adapters (**as of January 30, 2025**):

- [Phi4.Turn.R1Distill-Lora1](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill-Lora1)  
- [Phi4.Turn.R1Distill-Lora2](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill-Lora2)  
- [Phi4.Turn.R1Distill-Lora3](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill-Lora3)  
- [Phi4.Turn.R1Distill-Lora4](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill-Lora4)  
- [Phi4.Turn.R1Distill-Lora5](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill-Lora5)  
- [Phi4.Turn.R1Distill-Lora6](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill-Lora6)  
- [Phi4.Turn.R1Distill-Lora7](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill-Lora7)  
- [Phi4.Turn.R1Distill-Lora8](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill-Lora8)  

---

## GGUF Full & Quantized Models  
To facilitate broader testing and real-world inference, **GGUF Full and Quantized versions** have been provided for evaluation on **Open WebUI** and other LLM interfaces.

### **Version 1**
- [Phi4.Turn.R1Distill.Q8_0](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill.Q8_0)  
- [Phi4.Turn.R1Distill.Q4_k](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill.Q4_k)  
- [Phi4.Turn.R1Distill.16bit](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill.16bit)  

### **Version 1.1**
- [Phi4.Turn.R1Distill_v1.1_Q4_k](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill_v1.1_Q4_k)  

### **Version 1.2**
- [Phi4.Turn.R1Distill_v1.2_Q4_k](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill_v1.2_Q4_k)  

### **Version 1.3**
- [Phi4.Turn.R1Distill_v1.3_Q4_k-GGUF](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill_v1.3_Q4_k-GGUF)  

### **Version 1.4**
- [Phi4.Turn.R1Distill_v1.4_Q4_k-GGUF](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill_v1.4_Q4_k-GGUF)  

### **Version 1.5**
- [Phi4.Turn.R1Distill_v1.5_Q4_k-GGUF](https://huggingface.co/Quazim0t0/Phi4.Turn.R1Distill_v1.5_Q4_k-GGUF)  

---

## Usage  

### **Loading LoRA Adapters with `transformers` and `peft`**
To load and apply the LoRA adapters on Phi-4, use the following approach:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "microsoft/Phi-4"
lora_adapter = "Quazim0t0/Phi4.Turn.R1Distill-Lora1"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, lora_adapter)

model.eval()
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/Quazim0t0__Phi4.Turn.R1Distill_v1.5.1-Tensors-details)

|      Metric       |Value|
|-------------------|----:|
|Avg.               |22.67|
|IFEval (0-Shot)    |29.95|
|BBH (3-Shot)       |49.22|
|MATH Lvl 5 (4-Shot)| 1.59|
|GPQA (0-shot)      | 2.46|
|MuSR (0-shot)      | 7.04|
|MMLU-PRO (5-shot)  |45.75|