suayptalha/Turkish-Gemma-9b-T1-1bit

Turkish-Gemma-9b-T1 NanoQuant 1-bit

This repository contains a NanoQuant-compressed checkpoint derived from ytu-ce-cosmos/Turkish-Gemma-9b-T1.

Quantization details

• Method: NanoQuant
• Target linear-weight bitrate: 1.0 bit
• Sequence length: 2048
• Calibration samples: 64
• English calibration samples: 32
• Turkish calibration samples: 32
• Non-factorized epochs: 2
• Factorized epochs: 2
• ADMM outer iterations: 100
• Seed: 42
• Quantization hardware: NVIDIA A100 80 GB

Model-level knowledge-distillation tuning was performed after layer compression. The tuning loss decreased from 1.9244 to 1.7136 over 8 reported epochs.

Important

This file is a NanoQuant checkpoint. It is not a GGUF file and is not a standard Transformers save_pretrained checkpoint.

It requires the NanoQuant code and compatible CUDA kernels for loading and inference. It cannot be loaded directly with llama.cpp, Ollama, vLLM, or AutoModelForCausalLM.from_pretrained().

Files

• Turkish-Gemma-9b-T1-NanoQuant-1bit-en32-tr32.pt: compressed checkpoint

Base-model license

Use of this checkpoint remains subject to the license and usage terms of the original Gemma-based model.