license: llama2
datasets:
- uonlp/CulturaX
language:
- tr
- en
pipeline_tag: text-generation
metrics:
- accuracy
- bleu
Commencis-LLM
Commencis LLM is a generative model based on the Mistral 7B model. The base model adapts Mistral 7B to Turkish Banking specifically by training on a diverse dataset obtained through various methods, encompassing general Turkish and banking data.
Model Description
- Developed by: Commencis
- Language(s): Turkish
- Finetuned from model: Mistral 7B
- Input: Model input text only
- Output: Model generates text only
- Blog Post:
Training Details
Alignment phase consists of two stages: supervised fine-tuning (SFT) and Reward Modeling with Reinforcement learning from human feedback (RLHF).
The SFT phase was done on the a mixture of synthetic datasets generated from comprehensive banking dictionary data, synthetic datasets generated from banking-based domain and sub-domain headings, and derived from the CulturaX Turkish dataset by filtering. It was trained with three epochs. We used a learning rate 2e-5, lora rank 64 and maximum sequence length 1024 tokens.
Usage
Suggested Inference Parameters
- Temperature: 0.5
- Repetition penalty: 1.0
- Top-p: 0.9
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
class TextGenerationAssistant:
def __init__(self, model_id:str):
self.tokenizer = AutoTokenizer.from_pretrained(model_id)
self.model = AutoModelForCausalLM.from_pretrained(model_id, device_map='auto', load_in_8bit=True)
self.pipe = pipeline("text-generation",
model=self.model,
tokenizer=self.tokenizer,
device_map="auto",
max_new_tokens=1024,
return_full_text=True,
repetition_penalty=1.0
)
self.sampling_params = dict(do_sample=True, temperature=0.5, top_k=50, top_p=0.9)
self.SYSTEM_PROMPT = "Sen yardımcı bir asistansın. Sana verilen talimat ve girdilere en uygun cevapları üreteceksin. \n\n\n"
def format_prompt(self, user_input):
return "[INST] " + self.SYSTEM_PROMPT + user_input + " [/INST]"
def generate_response(self, user_query):
prompt = self.format_prompt(user_query)
outputs = self.pipe(prompt, **self.sampling_params)
return outputs[0]["generated_text"].split("[/INST]")[-1]
assistant = TextGenerationAssistant(model_id="Commencis/Commencis-LLM")
# Enter your query here.
user_query = "Faiz oranları yükseldiğinde kredilerim nasıl etkilenir?"
response = assistant.generate_response(user_query)
print(response)
Chat Template
from transformers import AutoTokenizer
import transformers
import torch
model = "Commencis/Commencis-LLM"
messages = [{"role": "user", "content": "Faiz oranları yükseldiğinde kredilerim nasıl etkilenir?"}]
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=1024, do_sample=True, temperature=0.5, top_k=50, top_p=0.9)
print(outputs[0]["generated_text"])
Bias, Risks, and Limitations
Like all LLMs, Commencis-LLM has certain limitations:
- Hallucination: Model may sometimes generate responses that contain plausible-sounding but factually incorrect or irrelevant information.
- Code Switching: The model might unintentionally switch between languages or dialects within a single response, affecting the coherence and understandability of the output.
- Repetition: The Model may produce repetitive phrases or sentences, leading to less engaging and informative responses.
- Coding and Math: The model's performance in generating accurate code or solving complex mathematical problems may be limited.
- Toxicity: The model could inadvertently generate responses containing inappropriate or harmful content.