RuBit-Llama-63M

This model is a fine-tuned version of NousResearch/Llama-2-7b-hf on the darulm dataset. From darulm aphorisms, dramaturgy, history, humor, literature domains were sampled

Training on 2_125_871_104 tokens.

Inspired by abideen/Bitnet-Llama-70M

Model description

Sample inference code

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load a pretrained BitNet model
model = "igorktech/RuBit-LLama-63M"
tokenizer = AutoTokenizer.from_pretrained(model)
model = AutoModelForCausalLM.from_pretrained(model)

def convert_to_bitnet(model, copy_weights):
    for name, module in model.named_modules():
        # Replace linear layers with BitNet
        if isinstance(module, LlamaSdpaAttention) or isinstance(module, LlamaMLP):
            for child_name, child_module in module.named_children():
                if isinstance(child_module, nn.Linear):
                    bitlinear = BitLinear(child_module.in_features, child_module.out_features, child_module.bias is not None).to(device="cuda:0")
                    if copy_weights:
                        bitlinear.weight = child_module.weight
                        if child_module.bias is not None:
                            bitlinear.bias = child_module.bias
                    setattr(module, child_name, bitlinear)
        # Remove redundant input_layernorms
        elif isinstance(module, LlamaDecoderLayer):
            for child_name, child_module in module.named_children():
                if isinstance(child_module, LlamaRMSNorm) and child_name == "input_layernorm":
                    setattr(module, child_name, nn.Identity().to(device="cuda:0"))
              

convert_to_bitnet(model, copy_weights=True)
model.to(device="cuda:0")

prompt = "Привет"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
generate_ids = model.generate(inputs.input_ids, max_length=100)
tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0015
• train_batch_size: 64
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 0.1
• num_epochs: 2
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.40.0
• Pytorch 2.2.1+cu121
• Datasets 2.19.0
• Tokenizers 0.19.1