m-a-p/Amber-Reproduce-50.33B

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.