README.md

metadata

license: llama3
language:
  - tr
  - en
base_model: meta-llama/Meta-Llama-3-8B-Instruct
model-index:
  - name: MARS
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge TR v0.2
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc
            value: 46.08
            name: accuracy
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU TR v0.2
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 47.02
            name: accuracy
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA TR v0.2
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: acc
            name: accuracy
            value: 49.38
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande TR v0.2
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 53.71
            name: accuracy
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k TR v0.2
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 53.08
            name: accuracy
pipeline_tag: text-generation

MARS

by Curiosity Technology

MARS is the first iteration of Curiosity Technology models, based on Llama 3 8B.

We have trained MARS on in-house Turkish dataset, as well as several open-source datasets and their Turkish translations. It is our intention to release Turkish translations in near future for community to have their go on them.

MARS have been trained for 3 days on 4xA100.

Model Details

• Base Model: Meta Llama 3 8B Instruct
• Training Dataset: In-house & Translated Open Source Turkish Datasets
• Training Method: LoRA Fine Tuning

How to use

You can run conversational inference using the Transformers pipeline abstraction, or by leveraging the Auto classes with the generate() function. Let's see examples of both.

Transformers pipeline

import transformers
import torch

model_id = "curiositytech/MARS"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "Sen korsan gibi konuşan bir korsan chatbotsun!"},
    {"role": "user", "content": "Sen kimsin?"},
]

terminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = pipeline(
    messages,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
print(outputs[0]["generated_text"][-1])

Transformers AutoModelForCausalLM

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "curiositytech/MARS"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "Sen korsan gibi konuşan bir korsan chatbotsun!"},
    {"role": "user", "content": "Sen kimsin?"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))