|
--- |
|
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](https://huggingface.co/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 [Your Name] at [Your Email]. |
