---

Phi-3.5-OASST1

A fine-tuned version of Phi-3.5 Mini Instruct trained on the OpenAssistant (OASST1) dataset using LoRA and 4-bit quantization. The goal of this project is to improve conversational and instruction-following capabilities while keeping training efficient through parameter-efficient fine-tuning.

Model Details

Base Model

• microsoft/Phi-3.5-mini-instruct

Fine-Tuning Dataset

• OpenAssistant/oasst1

Training Method

• LoRA (Low-Rank Adaptation)
• 4-bit Quantization using BitsAndBytes
• PEFT (Parameter Efficient Fine-Tuning)

Training Configuration

Parameter	Value
Base Model	Phi-3.5 Mini Instruct
Dataset	OpenAssistant/oasst1
LoRA Rank (r)	4
LoRA Alpha	32
Learning Rate	2e-5
Quantization	4-bit BitsAndBytes
Hardware	NVIDIA T4 GPU

Training Results

Final training metrics:

Train Loss: 2.5009
Global Steps: 4000
Training Runtime: 4507 seconds
Samples per Second: 0.887
Steps per Second: 0.887
Total FLOPs: 4.58e+16

Intended Use

This model is suitable for:

• Conversational AI
• Instruction Following
• Question Answering
• Educational Assistants
• General Text Generation

Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "YOUR_USERNAME/phi-3.5-oasst1"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Explain the difference between machine learning and deep learning."

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    temperature=0.7
)

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

Limitations

• Performance depends on prompt quality.
• Responses should be reviewed before use in critical applications.

Training Procedure

The model was fine-tuned on OpenAssistant conversations using LoRA adapters with 4-bit quantization to reduce memory usage and training costs. This approach allows efficient adaptation of the base model while retaining most of its original capabilities.

Acknowledgements

• Microsoft for Phi-3.5 Mini Instruct
• OpenAssistant for the OASST1 dataset
• Hugging Face Transformers
• PEFT
• BitsAndBytes

Citation

@misc{phi35_oasst1_2026,
  title={Phi-3.5-OASST1},
  author={Abubakar},
  year={2026},
  note={Phi-3.5 Mini Instruct fine-tuned on OpenAssistant OASST1 using LoRA and 4-bit quantization}
}

phi-3.5-oasst1

This model is a fine-tuned version of microsoft/Phi-3.5-mini-instruct on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 2.4659

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear

Training results

Training Loss	Epoch	Step	Validation Loss
2.4537	1.0	4000	2.4659

language:

• en license: mit base_model: microsoft/Phi-3.5-mini-instruct tags:
• phi-3
• lora
• qlora
• transformers
• conversational-ai
• instruction-tuning
• openassistant pipeline_tag: text-generation

---

Framework versions

• PEFT 0.19.1
• Transformers 5.12.0
• Pytorch 2.11.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2