Adapting LLMs to Hebrew: Unveiling DictaLM 2.0 with Enhanced Vocabulary and Instruction Capabilities
The DictaLM-2.0-Instruct Large Language Model (LLM) is an instruct fine-tuned version of the DictaLM-2.0 generative model using a variety of conversation datasets.
For full details of this model please read our release blog post or the technical report.
This is the instruct-tuned full-precision model designed for chat. You can try the model out on a live demo here.
You can view and access the full collection of base/instruct unquantized/quantized versions of DictaLM-2.0
here.
Instruction format
In order to leverage instruction fine-tuning, your prompt should be surrounded by [INST]
and [/INST]
tokens. The very first instruction should begin with a begin of sentence id. The next instructions should not. The assistant generation will be ended by the end-of-sentence token id.
E.g.
text = """<s>[INST] ืืืื ืจืืื ืืืื ืขืืื? [/INST]
ืืื, ืื ื ืื ืืืื ืืื ืืืคืืช ืืืฅ ืืืืื ืกืืื ืืจื. ืื ืืืกืืฃ ืืืืืง ืืช ืืืืืช ืื ืืื ื ืฉื ืืขื ืืืฆืืฅ ืืื ืื ืฉืื ื ืืืฉื ืืืืื!</s>[INST] ืืื ืืฉ ืื ืืชืืื ืื ืืืืื ื? [/INST]"
This format is available as a chat template via the apply_chat_template()
method:
Example Code
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
device = "cuda" # the device to load the model onto
model = AutoModelForCausalLM.from_pretrained("dicta-il/dictalm2.0-instruct", torch_dtype=torch.bfloat16, device_map=device)
tokenizer = AutoTokenizer.from_pretrained("dicta-il/dictalm2.0-instruct")
messages = [
{"role": "user", "content": "ืืืื ืจืืื ืืืื ืขืืื?"},
{"role": "assistant", "content": "ืืื, ืื ื ืื ืืืื ืืื ืืืคืืช ืืืฅ ืืืืื ืกืืื ืืจื. ืื ืืืกืืฃ ืืืืืง ืืช ืืืืืช ืื ืืื ื ืฉื ืืขื ืืืฆืืฅ ืืื ืื ืฉืื ื ืืืฉื ืืืืื!"},
{"role": "user", "content": "ืืื ืืฉ ืื ืืชืืื ืื ืืืืื ื?"}
]
encoded = tokenizer.apply_chat_template(messages, return_tensors="pt").to(device)
generated_ids = model.generate(encoded, max_new_tokens=50, do_sample=True)
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])
# <s> [INST] ืืืื ืจืืื ืืืื ืขืืื? [/INST]
# ืืื, ืื ื ืื ืืืื ืืื ืืืคืืช ืืืฅ ืืืืื ืกืืื ืืจื. ืื ืืืกืืฃ ืืืืืง ืืช ืืืืืช ืื ืืื ื ืฉื ืืขื ืืืฆืืฅ ืืื ืื ืฉืื ื ืืืฉื ืืืืื!</s> [INST] ืืื ืืฉ ืื ืืชืืื ืื ืืืืื ื? [/INST]
# ืืื, ืื ื ืืชืืื ืืกืืกื ืืงื ืืืื ืช ืืืื ื ืืืชื!
#
# ืืจืืืืื:
# - 2 ืืืืื ืื ืืืืืื
# - 1 ืืฃ ืืืืฅ ืืื ืืื
# (it stopped early because we set max_new_tokens=50)
Model Architecture
DictaLM-2.0-Instruct follows the Zephyr-7B-beta recipe for fine-tuning an instruct model, with an extended instruct dataset for Hebrew.
Limitations
The DictaLM 2.0 Instruct model is a demonstration that the base model can be fine-tuned to achieve compelling performance. It does not have any moderation mechanisms. We're looking forward to engaging with the community on ways to make the model finely respect guardrails, allowing for deployment in environments requiring moderated outputs.
Citation
If you use this model, please cite:
@misc{shmidman2024adaptingllmshebrewunveiling,
title={Adapting LLMs to Hebrew: Unveiling DictaLM 2.0 with Enhanced Vocabulary and Instruction Capabilities},
author={Shaltiel Shmidman and Avi Shmidman and Amir DN Cohen and Moshe Koppel},
year={2024},
eprint={2407.07080},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.07080},
}
- Downloads last month
- 50,378