llama-3-sauce-v2-8B / README.md
nbeerbower's picture
Adding Evaluation Results (#1)
9044672 verified
metadata
license: llama3
library_name: transformers
tags:
  - experimental
base_model:
  - nbeerbower/llama-3-bophades-v1-8B
datasets:
  - jondurbin/gutenberg-dpo-v0.1
  - jondurbin/truthy-dpo-v0.1
  - flammenai/FlameMix-DPO-v1
model-index:
  - name: llama-3-sauce-v2-8B
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 65.61
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=nbeerbower/llama-3-sauce-v2-8B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 83.11
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=nbeerbower/llama-3-sauce-v2-8B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 67.98
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=nbeerbower/llama-3-sauce-v2-8B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 56.39
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=nbeerbower/llama-3-sauce-v2-8B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 76.72
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=nbeerbower/llama-3-sauce-v2-8B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 72.48
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=nbeerbower/llama-3-sauce-v2-8B
          name: Open LLM Leaderboard

llama-3-sauce-v2-8B

This model is based on Llama-3-8b, and is governed by META LLAMA 3 COMMUNITY LICENSE AGREEMENT

This is a bad finetune on nbeerbower/llama-3-spicy-abliterated-stella-8B using various DPO sets.

Chat Format

Please use the ChatML format or you may experience poor results.

<|im_start|>system
{System Prompt Here!}<|im_end|>
<|im_start|>assistant
{Message from AI}<|im_end|>
<|im_start|>user
{Message from User}<|im_end|>

Method

Finetuned using an A100 on Google Colab.

Fine-tune a Mistral-7b model with Direct Preference Optimization - Maxime Labonne

Configuration

Dataset preparation:

def chatml_format(example):
    # Format system
    system = ""
    if example.get('system') and len(example['system']) > 0:
        systemMessage = example['system']
        system = "<|im_start|>system\n" + systemMessage + "<|im_end|>\n"

    # Format instruction
    prompt = "<|im_start|>user\n" + example['prompt'] + "<|im_end|>\n<|im_start|>assistant\n"

    # Format chosen answer
    chosen = example['chosen'] + "<|im_end|>\n"

    # Format rejected answer
    rejected = example['rejected'] + "<|im_end|>\n"

    return {
        "prompt": system + prompt,
        "chosen": chosen,
        "rejected": rejected,
    }

# Array of datasets to concat
ds = [
    "jondurbin/truthy-dpo-v0.1",
    "jondurbin/gutenberg-dpo-v0.1",
    "flammenai/FlameMix-DPO-v1"
]

# load_dataset and combine all
loaded_datasets = [load_dataset(dataset_name, split='train') for dataset_name in ds]
dataset = concatenate_datasets(loaded_datasets)

# Save columns
original_columns = dataset.column_names

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"

# Format dataset
dataset = dataset.map(
    chatml_format,
    remove_columns=original_columns
)

LoRA, model, and training settings:

# LoRA configuration
peft_config = LoraConfig(
    r=16,
    lora_alpha=16,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
)
# Model to fine-tune
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    load_in_4bit=True
)
model.config.use_cache = False
# Reference model
ref_model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    load_in_4bit=True
)
# Training arguments
training_args = TrainingArguments(
    per_device_train_batch_size=1,
    gradient_accumulation_steps=1,
    gradient_checkpointing=True,
    learning_rate=3e-5,
    lr_scheduler_type="cosine",
    max_steps=4000,
    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,
    force_use_ref_model=True
)
# Fine-tune model with DPO
dpo_trainer.train()

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 70.38
AI2 Reasoning Challenge (25-Shot) 65.61
HellaSwag (10-Shot) 83.11
MMLU (5-Shot) 67.98
TruthfulQA (0-shot) 56.39
Winogrande (5-shot) 76.72
GSM8k (5-shot) 72.48