Quantization made by Richard Erkhov.

Qwen2-1.5B-ITA-Instruct - GGUF

Model creator: https://huggingface.co/e-palmisano/
Original model: https://huggingface.co/e-palmisano/Qwen2-1.5B-ITA-Instruct/

Name	Quant method	Size
Qwen2-1.5B-ITA-Instruct.Q2_K.gguf	Q2_K	0.63GB
Qwen2-1.5B-ITA-Instruct.IQ3_XS.gguf	IQ3_XS	0.68GB
Qwen2-1.5B-ITA-Instruct.IQ3_S.gguf	IQ3_S	0.71GB
Qwen2-1.5B-ITA-Instruct.Q3_K_S.gguf	Q3_K_S	0.71GB
Qwen2-1.5B-ITA-Instruct.IQ3_M.gguf	IQ3_M	0.72GB
Qwen2-1.5B-ITA-Instruct.Q3_K.gguf	Q3_K	0.77GB
Qwen2-1.5B-ITA-Instruct.Q3_K_M.gguf	Q3_K_M	0.77GB
Qwen2-1.5B-ITA-Instruct.Q3_K_L.gguf	Q3_K_L	0.82GB
Qwen2-1.5B-ITA-Instruct.IQ4_XS.gguf	IQ4_XS	0.84GB
Qwen2-1.5B-ITA-Instruct.Q4_0.gguf	Q4_0	0.87GB
Qwen2-1.5B-ITA-Instruct.IQ4_NL.gguf	IQ4_NL	0.88GB
Qwen2-1.5B-ITA-Instruct.Q4_K_S.gguf	Q4_K_S	0.88GB
Qwen2-1.5B-ITA-Instruct.Q4_K.gguf	Q4_K	0.92GB
Qwen2-1.5B-ITA-Instruct.Q4_K_M.gguf	Q4_K_M	0.92GB
Qwen2-1.5B-ITA-Instruct.Q4_1.gguf	Q4_1	0.95GB
Qwen2-1.5B-ITA-Instruct.Q5_0.gguf	Q5_0	1.02GB
Qwen2-1.5B-ITA-Instruct.Q5_K_S.gguf	Q5_K_S	1.02GB
Qwen2-1.5B-ITA-Instruct.Q5_K.gguf	Q5_K	1.05GB
Qwen2-1.5B-ITA-Instruct.Q5_K_M.gguf	Q5_K_M	1.05GB
Qwen2-1.5B-ITA-Instruct.Q5_1.gguf	Q5_1	1.1GB
Qwen2-1.5B-ITA-Instruct.Q6_K.gguf	Q6_K	1.19GB
Qwen2-1.5B-ITA-Instruct.Q8_0.gguf	Q8_0	1.53GB

Original model description:

license: apache-2.0 datasets: - gsarti/clean_mc4_it - FreedomIntelligence/alpaca-gpt4-italian language: - it - en

This model has been fine-tuned with the continuous pretraining mode of Unsloth on the gsarti/clean_mc4_it dataset (only 100k rows) to improve the Italian language. The second fine-tuning was performed on the instructed dataset FreedomIntelligence/alpaca-gpt4-italian.

Uploaded model

Developed by: e-palmisano
License: apache-2.0
Finetuned from model : unsloth/Qwen2-1.5B-Instruct-bnb-4bit

Evaluation

For a detailed comparison of model performance, check out the Leaderboard for Italian Language Models.

Here's a breakdown of the performance metrics:

Metric	hellaswag_it acc_norm	arc_it acc_norm	m_mmlu_it 5-shot acc	Average
Accuracy Normalized	48.05	32.68	46.89	42.57

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.