akhooli / llama31ft

This Model (toy Arabic classical poetry llm)

This is a partially (one epoch, subset of Arabic classical poetry dataset) fine tuned Llama 3.1 8B LLM for poetry generation. It is based on a 10% of 1 epoch continued pretraining of the Llama 3.1 8B LLM . Training was done on 200k articles from Arabic Wikipedia 2023 with article lengh in the range 128 - 8192 words (not tokens).
This is just a proof of concept demo and should never be used for production. It is also not aligned and is likely to produce strange and unaccepted content. Only the adapter is available (along with other config files). To use it, you can either install Unsloth or use the HuggingFace PEFT API.
See installation instructions at the Unsloth's link below (only one GPU).
See the LinkedIn Post
and X tweet

Here's a simple usage example (raw output) - and remember, it is a primitive toy model using freely available compute.

max_seq_length = 256
dtype = None
load_in_4bit = True
​
alpaca_prompt = """
أدناه تعليمة تصف مهمة مقترنة بمدخلات تضيف سياق إن وجدت. اكتب إجابة تتناسب مع التعليمة والمدخلات مع الحفاظ على القيم واﻵداب العامة.
​
### التعليمة:
{}
​
### المدخلات:
{}
​
### اﻹجابة:
{}"""
​
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
        model_name = "akhooli/llama31ft", 
        max_seq_length = max_seq_length,
        dtype = dtype,
        load_in_4bit = load_in_4bit,
    )
model = FastLanguageModel.for_inference(model) 

inputs = tokenizer(
[
    alpaca_prompt.format(
        "اكتب قصيدة شعرية قصيرة", # instruction
        "بحر البسيط", # input
        "", # output - leave this blank for generation!
    )
], return_tensors = "pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens = 256, use_cache = True,temperature=0.95)
r = tokenizer.batch_decode(outputs)
from pprint import pprint
pprint(r)

Uploaded model

• Developed by: akhooli
• License: apache-2.0
• Finetuned from model : akhooli/llama31pretrained2

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