phi-1bee5 / README.md
pszemraj's picture
Update README.md
ec1f4cb
metadata
license: other
base_model: microsoft/phi-1_5
tags:
  - bees
  - honey
  - bzz
metrics:
  - accuracy
datasets:
  - BEE-spoke-data/bees-internal
language:
  - en
pipeline_tag: text-generation

phi-1bee5 🐝

Where Code Meets Beekeeping: An Unbeelievable Synergy!

Open In Colab

Have you ever found yourself in the depths of a debugging session and thought, "I wish I could be basking in the glory of a blooming beehive right now"? Or maybe you've been donning your beekeeping suit, puffing on your smoker, and longed for the sweet aroma of freshly written code?

Well, brace yourselves, hive-minded humans and syntax-loving sapiens, for phi-1bee5, a groundbreaking transformer model that's here to disrupt your apiary and your IDE!

Details

This model is a fine-tuned version of microsoft/phi-1_5 on the BEE-spoke-data/bees-internal dataset.

It achieves the following results on the evaluation set:

  • Loss: 2.6982
  • Accuracy: 0.4597

Usage

load model:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
# !pip install -U -q transformers accelerate einops

checkpoint = "BEE-spoke-data/phi-1bee5"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(
    checkpoint, 
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True
)

Run inference:

prompt = "Today was an amazing day because"
inputs = tokenizer(prompt, return_tensors="pt", return_attention_mask=False).to(
    model.device
)

outputs = model.generate(
    **inputs, do_sample=True, max_new_tokens=128, epsilon_cutoff=7e-4
)
result = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
print(result)
# output will probably contain a story/info about bees

Intended Uses:

  1. Educational Edification: Are you a coding novice with a budding interest in beekeeping? Or perhaps a seasoned developer whose curiosity has been piqued by the buzzing in your backyard? phi-1bee5 aims to serve as a fun, informative bridge between these two worlds.
  2. Casual Queries: This model can generate code examples and beekeeping tips. It's perfect for those late-night coding sessions when you feel like taking a virtual stroll through an apiary.
  3. Academic & Research Insights: Interested in interdisciplinary studies that explore the intersection of technology and ecology? phi-1bee5 might offer some amusing, if not entirely accurate, insights.

Limitations:

  1. Not a beekeeping expert: For the love of all things hexagonal, please do not use phi-1bee5 to make serious beekeeping decisions. While our model is well read in the beekeeping literature, it lacks the practical experience and nuanced understanding that professional beekeepers possess.
  2. Licensing: This model is derived from a base model under the Microsoft Research License. Any use must comply with the terms of that license.
  3. Infallibility: Like any machine learning model, phi-1bee5 can make mistakes. Always double check the code and bee facts before using it in production or in your hive.
  4. Ethical Constraints: This model may not be used for illegal or unethical activities, including but not limited to terrorism, harassment, or spreading disinformation.

Training procedure

While the full dataset is not yet complete and therefore not yet released for "safety reasons", you can check out a preliminary sample at: bees-v0

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0001
  • train_batch_size: 1
  • eval_batch_size: 2
  • gradient_accumulation_steps: 32
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.995) and epsilon=1e-08
  • lr_scheduler_type: cosine
  • lr_scheduler_warmup_ratio: 0.03
  • num_epochs: 2.0