IMG7/Sarvam30bCompression

Model Card: Compressed Sarvam-30b Submission

Model Details

• Model name: Sarvam-30b Compressed Submission
• Base model: Sarvam-30b
• Track: Resilient AI Challenge – Sarvam category
• Task type: Text-to-text
• Inference backend: vLLM or llama.cpp
• Compression type: architecture-preserving inference optimization

Intended Use

This model is intended exclusively for challenge evaluation in the Sarvam-30b category. It is designed to preserve the original functional behavior of the base model while improving inference efficiency.

Training / Adaptation

No architectural redesign was introduced. The submission focuses on compression and inference-time optimization while keeping the original model structure intact.

Evaluation Considerations

The submission was prepared to balance two goals:

1. maintain quality relative to the Sarvam-30b baseline;
2. reduce energy consumption during inference.

Repository Contents

• model weights
• serving configuration file
• README describing the methodology

Access

This repository is shared with the organizing team for evaluation purposes.