mncai_-_mistral-7b-dpo-v6-4bits

Quantization made by Richard Erkhov.

Github

Discord

Request more models

mistral-7b-dpo-v6 - bnb 4bits

• Model creator: https://huggingface.co/mncai/
• Original model: https://huggingface.co/mncai/mistral-7b-dpo-v6/

Original model description:

license: apache-2.0 datasets:

• Intel/orca_dpo_pairs language:
• en

---

Model Card for mncai/mistral-7b-dpo-v6

Introduction of MindsAndCompany

https://mnc.ai/

We create various AI models and develop solutions that can be applied to businesses. And as for generative AI, we are developing products like Code Assistant, TOD Chatbot, LLMOps, and are in the process of developing Enterprise AGI (Artificial General Intelligence).

Model Summary

based mistral-7b, dpo tuned.

Detail

first step ties merge.

models:
  - model: AIDC-ai-business/Marcoroni-7B-v3
    # no parameters necessary for base model
  - model: GreenNode/GreenNodeLM-7B-v1olet # psmathur/orca_mini_v3_13b
    parameters:
      density: [1, 0.7, 0.1] # density gradient
      weight: 1.0
  - model: viethq188/LeoScorpius-7B-Chat-DPO
    parameters:
      density: 0.5
      weight: [0, 0.3, 0.7, 1] # weight gradient
  - model: mncai/mistral-7b-dpo-v5
    parameters:
      density: 0.33
      weight:
        - filter: mlp
          value: 0.5
        - value: 0
merge_method: ties
base_model: AIDC-ai-business/Marcoroni-7B-v3
parameters:
  normalize: true
  int8_mask: true
dtype: float16

second step dpo.

# Training arguments
training_args = TrainingArguments(
    per_device_train_batch_size=5,
    gradient_accumulation_steps=4,
    gradient_checkpointing=True,
    learning_rate=5e-6,
    lr_scheduler_type="cosine",
    max_steps=1000,
    save_strategy="no",
    logging_steps=1,
    output_dir=new_model,
    optim="paged_adamw_32bit",
    warmup_steps=100,
    bf16=True,
    report_to="wandb",
)

# Create DPO trainer
dpo_trainer = DPOTrainer(
    model,
    ref_model,
    args=training_args,
    train_dataset=dataset,
    tokenizer=tokenizer,
    # peft_config=peft_config,
    beta=0.1,
    max_prompt_length=1024,
    max_length=2048,
)

# Fine-tune model with DPO
dpo_trainer.train()

How to Use

Here give some examples of how to use our model.

from transformers import AutoConfig, AutoModel, AutoTokenizer
import transformers
import torch
hf_model = 'mncai/mistral-7b-dpo-v6' 
message = "<|user|>\n두 개의 구가 있는데 각각 지름이 1, 2일때 구의 부피는 몇배 차이가 나지? 설명도 같이 해줘.\n<|assistant|>\n"

sequences = pipeline(
    message,
    do_sample=True,
    top_k=10,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
    max_length=2048,
)
for seq in sequences:
    print(f"Result: {seq['generated_text']}")

Warnings

Currently, the leaderboard is overfitted. It is inevitable because, unlike Kaggle, where there's private scoring followed by the end of the competition, here the scores are continuously open. Even among my models, some received lower scores in internal data evaluations. mncai/agiin-13.6B-v0.1 > mncai/agiin-11.1B-v0.1 > mncai/mistral-7b-dpo-v6. However, on the leaderboard, mncai/mistral-7b-dpo-v6 has the highest score. When choosing a model to use on the open LLM leaderboard, it would be best to evaluate with your own private dataset that is not publicly available.

Detect-Pretrain-Code-Contamination Result Share

use https://github.com/Mihaiii/detect-pretrain-code-contamination

DATASET=truthful_qa python src/run.py --target_model mncai/mistral-7b-dpo-v6  --data $DATASET --output_dir out/$DATASET --ratio_gen 0.4

result < 0.1, %: 0.76

Contact

If you have any questions, please raise an issue or contact us at [email protected]