metadata

language: code
tags:
  - code
  - gpt2
  - generation
datasets:
  - giulio98/xlcost-single-prompt
widget:
  - text: |
      '''
      function to add two numbers
      '''
      ###
    example_title: add two numbers
model-index:
  - name: codegen-350M-multi-xlcost
    results:
      - task:
          name: Code Generation
          type: code-generation
        dataset:
          name: XLCost
          type: code_eval_outputs
        metrics:
          - name: pass@1
            type: code_eval_outputs
            value: 3.325
          - name: pass@10
            type: code_eval_outputs
            value: 15
          - name: codebleu
            type: codebleu
            value: 20.18191

CodeGen-350M-multi-xlcost-v2

CodeGen-350M-multi-xlcost is a CodeGen model fine-tuned on the Python split of XLCost dataset using Deepspeed.

Usage

You can load the CodeGen-350M-multi-xlcost-v2 model and tokenizer directly in transformers:

from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("giulio98/codegen-350M-multi-xlcost-v2")
model = AutoModelForCausalLM.from_pretrained("giulio98/codegen-350M-multi-xlcost-v2")

text = tokenizer.eos_token + "\'\'\'\n" + "function to add two numbers" + "\n\'\'\'\n" + "###\n"
input_ids = tokenizer(text, return_tensors="pt").input_ids

generated_ids = model.generate(input_ids, max_length=128)
print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))

Output:

'''
function to add two numbers 
'''
###
def add(a, b):
    return a + b

Training

The model was finetuned on XLCost-single-prompt, an improved version of the original XLCost dataset xlcost-text-to-code. Below the hyperparameters.

Hyperparameter	value
Per device train batch size	16
Context size	1024
Training steps	259
Gradient accumulation	2
Gradient checkpointing	True
Learning rate	1.8e-05
Weight decay	0.1
Warmup steps	35
Schedule	linear
zero stage	2

Below the deepspeed configuration

{
  "fp16": {
    "enabled": true,
    "loss_scale": 0,
    "loss_scale_window": 1000,
    "initial_scale_power": 16,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": 0.000018,
      "betas": [
        0.9,
        0.999
      ],
      "eps": 1e-8,
      "weight_decay": 0.1
    }
  },
  "scheduler": {
    "type": "WarmupLR",
    "params": {
      "warmup_min_lr": 0,
      "warmup_max_lr": 0.000018,
      "warmup_num_steps": 35
    }
  },
  "zero_optimization": {
    "stage": 2,
    "offload_optimizer": {
      "device": "cpu",
      "pin_memory": false
    },
    "allgather_partitions": true,
    "allgather_bucket_size": 200000000,
    "overlap_comm": true,
    "reduce_scatter": true,
    "reduce_bucket_size": 200000000,
    "contiguous_gradients": true
  },
  "gradient_accumulation_steps": 2,
  "train_batch_size": 32,
  "train_micro_batch_size_per_gpu": 16,
  "gradient_clipping": 1,
  "wall_clock_breakdown": false
}

The training was executed on 1 x V100 (16GB) GPU for 28min 50sec

Performance

We evaluated the model on the first 400 samples of XLCOST's XLCost-single-prompt test split and comparing the outputs of the generated codes with respect to the expected output using pass@k metric.

Metric	codegen-350M-multi-xlcost-v2	codegen-350M-multi-xlcost	codegen-350M-mono(zero-shot)	codegen-350M-mono (one-shot)	codegen-350M-mono(few-shot)
pass@1	3.325%	3.70%	0.4%	0.35%	0.48%
pass@10	15%	14.5%	3.5%	3 %	3.75%
CodeBLEU	20.18%	None	15.15%	19.42 %	20.27%

The pass@k metric tells the probability that at least one out of k generations passes the tests.

Citations

@article{Nijkamp2022ACP,
  title={A Conversational Paradigm for Program Synthesis},
  author={Nijkamp, Erik and Pang, Bo and Hayashi, Hiroaki and Tu, Lifu and Wang, Huan and Zhou, Yingbo and Savarese, Silvio and Xiong, Caiming},
  journal={arXiv preprint},
  year={2022}
}