The datagemma-2b is a model designated for data science code generation from natural language instruction. It is fine-tuned from codegemma-2b model. Fine tuning was performed on the ed001/ds-coder-instruct-v2 dataset which is constructed by filtering publicly available datasets on HuggingFace.

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model = AutoModelForCausalLM.from_pretrained(
    "ed001/datagemma-2b",
    low_cpu_mem_usage=True
).cuda()

# Reload tokenizer to save it
tokenizer = AutoTokenizer.from_pretrained("ed001/datagemma-2b", trust_remote_code=True)
tokenizer.padding_side = "right"

prompt_template =  "### Question: {}\n ### Answer: "
generation_config = GenerationConfig(max_new_tokens=512, top_p=0.5, do_sample=True, repetition_penalty=1)
prompt = "How can I profile speed of my neural network using PyTorch?"
input = tokenizer(prompt_template.format(prompt), return_tensors="pt").to(model.device)["input_ids"]

print(tokenizer.decode(model.generate(input, generation_config=generation_config)[0]))

Training Details

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
target_modules: q, k, v, o, gate_proj, down_proj, up_proj
weight_decay: 0
optmizer: paged_adamw_8bit
lr: 1e-4
lr_scheduler: cosine
max_seq_len: 1536
batch_size: 1 grad_acc: 4 max_grad_norm: 0.5
warmup_ratio: 0.05
num_epochs: 1

Contact

GitHub: Ea0011