base_model: microsoft/Phi-3-mini-4k-instruct
library_name: peft
pipeline_tag: text-generation
tags:
- base_model:adapter:microsoft/Phi-3-mini-4k-instruct
- lora
- transformers
license: apache-2.0
datasets:
- teknium/OpenHermes-2.5
- Magpie-Align/Magpie-Phi3-Pro-300K-Filtered
language:
- en
Model Card for Model ID
MicroAtlas-V1 is a general purpose model trained on both the teknium/OpenHermes-2.5 dataset and the Magpie-Align/Phi3-Pro-300K-Filtered dataset and designed to provide speed, efficiency, and intelligence while still being relatively small.
the .gguf version of this model has only 3.15 M parameters, making it extremely small.
Model Details
OpenHermes dataset:
1 Epoch
8 Batch Size
1 Gradient Accumulation
5e-5 LR
16 LoRa r
32 LoRa Alpha
300 Warmup steps
500 Eval steps
Trained only on Attention layers.
Magpie dataset:
1 Epoch
16 Batch Size
1 Gradient Accumulation
1e-4 LR
16 LoRa r
32 LoRa Alpha
150 Warmup steps
500 Eval steps
Trained with Gate,Up, and Down layers.
Model Description
This model excels at creating bullet point formatting.
- Developed by: Turtle170 (anonymous
- Language(s) (NLP): English
- License: apache-2.0
- Finetuned from model : Phi-3-Mini-4k-Instruct with turtle170/Phi-3-Mini-OpenHermes-V1 adapters
Direct Use
For direct use, the easiest method is to just download the .gguf file from turtle170/Phi-3-Mini-OpenHermes-Magpie-V1-F16-GGUF and loading it into llama.cpp or Ollama.
Out-of-Scope Use
Users of this model need only adhere to the Microsoft Phi-3 Terms of use, and you are solely responsible for any misuse of this model, as according to Section 7 and 8 of the apache-2.0 licence
Bias, Risks, and Limitations
As this model was trained on a small base model, and only exposed to 2 50k example datasets, so you should not expect much from it. However, this model is smart for its size.
Recommendations
This model
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
Training Details
Stated above.
Training Data
teknium/OpenHermes-2.5 dataset and the Magpie-Align/Phi3-Pro-300K-Filtered dataset.
Training Procedure
Training Hyperparameters
- Training regime: The OpenHermes run was on fp16 mixed precision, while the Magpie run was on fp32 mixed precison.
Speeds, Sizes, Times
The Magpie adapter is about 100-200 Mb.
Evaluation
The evaluation strategy was epochs, and the results were 0.4203 loss.
Metrics
1 Epoch --> fast while prevents overfitting.
16 Batch Size --> Helps squeeze every bit of intelligence.
1 Gradient Accumulation --> fast, while not crashing the model.
1e-4 LR --> helps prevent breaking the intelligence stored on the Hermes run.
16 LoRa r --> helps the model understand the harder examples in the Magpie run.
32 LoRa Alpha --> self-explanatory. Alpha = LoRa r x 2
150 Warmup steps -->fast, and since the starting loss was already 0.4
1500 Eval steps --> the loss had fluctuated between 0.4 and 0.6, and eval wastes time, so i chose it to only be 2 per run.
Results
eval loss: 0.4 Avg. train loss: 0.4
Environmental Impact
- Hardware Type: 2x NVIDIA Tesla T4s
- Hours used: 12
- Cloud Provider: Kaggle
- Compute Region: asia-east1
- Carbon Emitted: 0.47 kg
Model Architecture and Objective
To provide a smart model while keeping the size small.
Framework versions
- PEFT 0.17.1