metadata
library_name: transformers
tags:
- unsloth
license: apache-2.0
datasets:
- lighteval/MATH-Hard
language:
- en
- th
- pt
- es
- de
- fr
- it
- hi
base_model:
- meta-llama/Llama-3.2-3B-Instruct
metrics:
- accuracy
This version of Komodo is a Llama-3.2-3B-Instruct finetuned model on lighteval/MATH-Hard dataset to increase math performance of the base model.
This model is 4bit-quantized. You should import it 8bit if you want to use 3B parameters! Make sure you installed 'bitsandbytes' library before import.
Example Usage:
tokenizer = AutoTokenizer.from_pretrained("suayptalha/Komodo-Llama-3.2-8B")
model = AutoModelForCausalLM.from_pretrained("suayptalha/Komodo-Llama-3.2-8B")
example_prompt = """Below is a math question and its solution:
Question: {}
Solution: {}"""
inputs = tokenizer(
[
example_prompt.format(
"", #Question here
"", #Solution here (for training)
)
], return_tensors = "pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens = 50, use_cache = True)
tokenizer.batch_decode(outputs)
