ahmadhehe/tinyllama-1.1b-sft-dolly15k
The ahmadhehe/tinyllama-1.1b-sft-dolly15k model is a 1.1 billion parameter TinyLlama variant, fine-tuned by Ahmad Murtaza and Simra Sheikh using Supervised Fine-Tuning (SFT) on the databricks/databricks-dolly-15k dataset. This model specializes in instruction following, demonstrating improved performance over its base model on a 10-prompt test set, achieving a BLEU-4 score of 2.4200 and a BERTScore F1 of 87.1100. It is designed for tasks requiring responses based on instructions, making it suitable for conversational agents or task-oriented dialogue systems.
Loading preview...
Model Overview
This model, ahmadhehe/tinyllama-1.1b-sft-dolly15k, is a 1.1 billion parameter language model based on the TinyLlama architecture. It was developed by Ahmad Murtaza and Simra Sheikh as part of an NLP course assignment, focusing on Supervised Fine-Tuning (SFT) using the databricks/databricks-dolly-15k dataset.
Key Capabilities
- Instruction Following: Fine-tuned specifically for instruction-based tasks, enabling it to generate responses aligned with given prompts.
- Improved Performance: Demonstrates enhanced performance over the base TinyLlama-1.1B model, achieving a BLEU-4 score of 2.4200 and a BERTScore F1 of 87.1100 on a 10-prompt test set, compared to the base model's 2.1400 and 85.4700 respectively.
- Efficient Fine-tuning: Utilizes LoRA (Low-Rank Adaptation) with a rank of 16 and alpha of 32, trained for 1 epoch with a learning rate of 0.0002, resulting in a validation loss of 1.7798.
Good For
- Instruction-tuned applications: Ideal for use cases where the model needs to follow specific instructions or generate responses based on prompts.
- Resource-constrained environments: As a 1.1 billion parameter model, it offers a balance between capability and computational efficiency.
- Educational and research projects: Serves as a practical example of an SFT pipeline on a smaller LLM, suitable for further experimentation and development.