Amber-Reproduce-Intermediate-CKPTs (The Fine Line)
Collection
23 items
•
Updated
YAML Metadata
Warning:
empty or missing yaml metadata in repo card
(https://huggingface.co/docs/hub/model-cards#model-card-metadata)
Architecture & Training Configuration:
Base Model Configuration: This variant is built upon the Llama2-7B configuration, ensuring a robust foundation that aligns with the latest advancements in model architecture.
Sequence Length Adaptation: Originally processed data for a sequence length of 2048 was detokenized and re-encoded to fit a sequence length of 4096. This step follows the preprocessing strategy of Megatron-LM, enhancing our model's capacity to understand and generate more complex sequences.
Batch Size & Token Management: We adopted a batch size capable of managing 4 million tokens, tailored to accommodate the increased sequence length and ensure efficient data processing.
Integration of GQA Technologies: To boost training efficiency, our configuration includes the integration of Gradient Quantization and Aggregation technologies. With 32 attention heads and a group size of 4, this feature significantly enhances the model's learning and processing capabilities.