Multilingual language models have many deployment challenges.

Can we engineer multilingual language models that not only match the prowess of their bulkier counterparts but do so while being more compact, quicker on their feet, and capable of handling massive data batches in real-time production environments. Is this a feat we can achieve?

Techniques:

• Pruning

  • Unstructured Pruning

  • Structured Pruning

  • Semi-Structured Pruning

  • Methods Used

    • SparseGPT | GitHub
    • ShortGPT | KLDBasedPruning & Perplexity Sensivities

• Knowledge Distillation

  • Hidden State-Based Distillation ~ DistillKit | GitHub
  • Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling
  • On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes
  • Minitron: Compact Language models via Pruning & Knowledge Distillation
  • DistiLLM: Towards Streamlined Distillation for Large Language Models

• Quantization

  • Quantization Aware Training (QAT)
  • Post Training Quantization (PTQ)
    • KV Cache Quantization
    • Weight & Activation Quantization

• Low-Rank Factorization

• Fine-Tuning | GitHub

Datasets:

Initial 7 datasets unified, having 6.62M rows which includes the following:

• Bangla_Alpaca_Orca : Bangle
• Urdu_Instruct_News_Article_Generation: Urdu
• Urdu_Instruct_News_Headline_Generation: Urdu
• Urdu_Instruct_News_Category_Classification: Urdu
• cidar: Arabic
• Six_Millions_Instruction_Dataset_For_Arabic_Llm_Ft: Arabic
• instructv3: English

Get in touch with the team:

• Mayank Bhaskar -> [email protected]
• Ahmad Anis -> [email protected]
• Drishti Sharma -> [email protected]
• Vishnu Vardhan -> [email protected]
• Yaya -> [email protected]
• Shayekh Bin Islam -> [email protected]