metadata

language:
  - en
license: apache-2.0
tags:
  - merge
  - mergekit
  - lazymergekit
  - meta-llama/Meta-Llama-3-8B
  - beratcmn/Llama-3-11.5B
  - text-generation-inference
  - transformers
  - unsloth
  - llama
  - trl
base_model: beratcmn/Llama-3-11.5B

Llama-3-11.5B

This model is a Proof of Concept. First 2 Llama-3-8B models has been merged using Mergekit and pre-training continued using QLora and Unsloth for 1000 samples from roneneldan/TinyStories. Loss still decreases each epoch so I believe this is a successful experiment where there is a lot of room to experiment.

Wandb Report

Llama-3-11.5B is a merge of the following models using LazyMergekit:

🧩 Configuration

slices:
  - sources:
    - model: meta-llama/Meta-Llama-3-8B
      layer_range: [0, 24]
  - sources:
    - model: meta-llama/Meta-Llama-3-8B
      layer_range: [8, 32]
merge_method: passthrough
dtype: bfloat16

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "beratcmn/Llama-3-11.5B"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Uploaded model

Developed by: beratcmn
License: apache-2.0
Finetuned from model : beratcmn/Llama-3-11.5B

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