This is kaitchup/Qwen2.5-1.5B-Minivoc-32k-v0.1a quantized with AutoRound (asymmetric quantization) and serialized with the GPTQ format in 4-bit. The model has been created, tested, and evaluated by The Kaitchup.

The Minivoc approach reduces the vocabulary size to save memory during inference and fine-tuning.

Details on the quantization process and how to use the model here: The Best Quantization Methods to Run Llama 3.1 on Your GPU

It is possible to fine-tune an adapter on top of it following the QLoRA methodology. More about this here: QLoRA with AutoRound: Cheaper and Better LLM Fine-tuning on Your GPU

I used these hyperparameters for quantization:

bits, group_size = 4, 128

autoround = AutoRound(model, tokenizer, nsamples=512, iters=1000, low_gpu_mem_usage=False, bits=bits, group_size=group_size)

autoround.quantize()
output_dir = "./tmp_autoround"
autoround.save_quantized(output_dir, format='auto_gptq', inplace=True)

Evaluation results (zero-shot evaluation with lm_eval):

arc_challenge, musr, gpqa, mmlu_pro, mmlu….png

Developed by: The Kaitchup
Language(s) (NLP): English
License: Apache 2.0 License

kaitchup/Qwen2.5-1.5B-Minivoc-32k-v0.1a-AutoRound-GPTQ-asym-4bit

Model Details

kaitchup/
Qwen2.5-1.5B-Minivoc-32k-v0.1a-AutoRound-GPTQ-asym-4bit