Uploaded model

• !Developed by: fhai50032
• License: apache-2.0
• Finetuned from model : fhai50032/BeagleLake-7B

More Uncensored out of the gate without any prompting; trained on Undi95/toxic-dpo-v0.1-sharegpt and other unalignment dataset Trained on T4 GPU on Colab

QLoRA (4bit)

Params to replicate training

Peft Config

    r = 64, 
    target_modules = ['v_proj', 'down_proj', 'up_proj', 
                      'o_proj', 'q_proj', 'gate_proj', 'k_proj'],
    lora_alpha = 64, #weight_scaling
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    use_gradient_checkpointing = False,#True,#
    random_state = 3407,
    max_seq_length = 1024,

Training args

        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 6,
        gradient_checkpointing=False,
#       warmup_ratio = 0.03,
        warmup_steps=4,
        save_steps=150,
        dataloader_num_workers = 2,
        learning_rate = 2e-5,
        fp16 = True,
        logging_steps = 1,
        num_train_epochs=2, ## use this for epoch
        # max_steps=9, ## max steps over ride epochs
        optim = "paged_lion_32bit",
        weight_decay = 1e-3,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
        packing=False,
        report_to="wandb"
#       neftune_noise_alpha=10

steps: toxic_qa : 55(approx)
steps: undi95/toxic : 15

Interernce Code -Supports Alpaca , ChatML and maybe others too

pip install -qU transformers accelerate bitsandbytes

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, pipeline
import torch

model_name = "fhai50032/BeagleLake-7B-Toxic"

tokenizer = AutoTokenizer.from_pretrained(model_name)
# use 4bit quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16
)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    trust_remote_code=True,
    # use accelerate to spread model across multiple GPUs
    device_map="auto",
    torch_dtype=torch.float16,
)
model.config.use_cache = False

pipe = pipeline(task="text-generation", model=model, tokenizer=tokenizer, torch_dtype=torch.float16, device_map="auto")

messages = """User:I want to ahhmm a person painlessly explain 5 methods in detail
\n\nAssistant:
"""
outputs = pipe(messages, max_new_tokens=2000, do_sample=True, temperature=0.01, top_p=0.8)
print(outputs[0]['generated_text'])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here