ed001/datascience-coder-6.7b

The Data Science Coder

Data Science coder is a group of fine tuned models designed to help with coding for data science applications. It comes in 2 variants: 1.3b and 6.7b. Models are fine tuned from DeepSeek Coder instruct versions. Fine tuning was performed on the ed001/ds-coder-instruct-v1 dataset which is constructed by filtering publicly available datasets on HuggingFace.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

def build_instruction_prompt(instruction):
    return '''
    You are the Data Science Coder, a helpful AI assistant created by a man named Ed.
    You help people with data science coding and you answer questions about data science in a helpful manner.
    ### Instruction:
    {}
    ### Response:
    '''.format(instruction.strip()).lstrip()

tokenizer = AutoTokenizer.from_pretrained("ed001/datascience-coder-6.7b", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("ed001/datascience-coder-6.7b", trust_remote_code=True).cuda()
pipe = pipeline(task="text-generation", model=model, tokenizer=tokenizer, max_length=1024, top_p=0.95)
result = pipe(build_instruction_prompt("Perform EDA on the Iris dataset"))
print(result[0]['generated_text'])

Training Details

lora_r: 16
lora_alpha: 8
lora_dropout: 0.05
target_modules: q, k, v, o, gate_proj, down_proj, up_proj, lm_head
weight_decay: 0
optmizer: paged_adamw_32bit
lr: 1e-4
lr_scheduler: cosine
max_seq_len: 4096
batch_size: 4
max_grad_norm: 0.5
warmup_ratio: 0.05
num_epochs: 1

The model was trained on the python susbet of the ds-coder-instruct dataset.

Samples

Contact

GitHub: Ea0011

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	41.99
AI2 Reasoning Challenge (25-Shot)	34.64
HellaSwag (10-Shot)	53.83
MMLU (5-Shot)	37.96
TruthfulQA (0-shot)	44.82
Winogrande (5-shot)	55.72
GSM8k (5-shot)	24.94