Kukedlc_-_NeuralExperiment-7b-MagicCoder-v7.5-4bits

Quantization made by Richard Erkhov.

Github

Discord

Request more models

NeuralExperiment-7b-MagicCoder-v7.5 - bnb 4bits

• Model creator: https://huggingface.co/Kukedlc/
• Original model: https://huggingface.co/Kukedlc/NeuralExperiment-7b-MagicCoder-v7.5/

Original model description:

license: apache-2.0 datasets:

• microsoft/orca-math-word-problems-200k
• ise-uiuc/Magicoder-Evol-Instruct-110K
• Vezora/Tested-22k-Python-Alpaca

---

Datacard for Custom Trained Model

• Base Model : Kukedlc/NeuralExperiment-7b-dare-ties

Model Description

This model is an experimental AI trained on three distinct datasets focusing on logical reasoning, mathematics, and programming. The training process involved fine-tuning from the last layer (31) backward with a gradually decreasing learning rate. The primary goal is to address and rectify the common 'INSTINST' bug observed in leaderboard models through targeted training on the latest layers.

Datasets Used for Training

• microsoft/orca-math-word-problems-200k: A large-scale dataset of mathematical word problems aimed at enhancing the model's numerical reasoning and problem-solving capabilities.
• ise-uiuc/Magicoder-Evol-Instruct-110K: A dataset designed to improve code generation and understanding, contributing to the model's programming language proficiency.
• sahil2801/CodeAlpaca-20k: A dataset focused on programming challenges to further refine the model's coding and logical reasoning skills.

Each dataset contributed 20,000 data points to the training process, ensuring a balanced representation of logic, mathematics, and programming tasks.

Training Environment

• The model was trained on Kaggle's free GPU environment, allowing for cost-effective fine-tuning and experimentation.
• Users interested in replicating or extending this training can find the Kaggle notebook in my profile or request it directly for collaborative purposes.

Preliminary Results

• The model shows promising results in solving logical puzzles and mathematical problems, especially those with misleading or non-obvious solutions that it initially struggled with.
• Ongoing experiments aim to quantify the impact of targeted training on the model's reasoning capabilities across different domains.

Invitation for Collaboration

• Feedback, suggestions, and collaborative efforts are highly encouraged to further refine and evaluate the model.
• If interested in contributing or experimenting with this model, please feel free to reach out or access the code directly from my Kaggle profile.

Contact Information

• For any inquiries, suggestions, or collaboration proposals, please contact me!

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "Kukedlc/NeuralExperiment-7b-MagicCoder-v7"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])