zredlined

NL2Code Workshop 2024

Powering the Shift from Code to Conversation talk - Keynote References

1. Benchmarks and Leaderboards

• HumanEval Benchmark (Code Generation) - Papers with Code
  An overview of the current state-of-the-art in code generation on the HumanEval benchmark.
• Announcing BigCodeBench-Hard, and More - Hugging Face
  Introduction to BigCodeBench-Hard, an enhanced challenge for code generation models.

2. Impact of Code in Pre-training

zredlined

prompt = f"""
Compare the quality of the following two instruction-response pairs:


Version 1 ({version1_name}):
Instruction: {version1['instruction']}
Context: {version1['context']}
Response: {version1['response']}

zredlined

CUDA_VISIBLE_DEVICES=0 python train.py
--gpu_ids 0 --dataroot ./datasets/ebike_data
--name locations_FastCUT --CUT_mode FastCUT --n_epochs 30

zredlined

import logging

from gretel_client import configure_session, ClientConfig, projects
from gretel_client.helpers import do_api_call

logging.getLogger().setLevel(logging.INFO)

def clean_projects():
    """ Delete any projects that are not shared to more than one person """
    for project in projects.search_projects():

zredlined

def plot_timeseries(filenames: list):
    """ Plot a list of timeseries CSVs to a single graph """
    combined_df = pd.DataFrame()
    
    for filename in filenames:
        df = pd.read_csv(filename)
        df['date'] = pd.to_datetime(df['date'])
        df['label'] = filename
        combined_df = pd.concat([combined_df, df])
        

zredlined

# Use the model to generate synthetic datasets

seed_data = json.loads(trends_df[seed_fields].to_json(orient="records"))

for dataset in range(5):
    model.generate(max_invalid=1e5,
                   num_proc=1, # disable paralellism when using seed data
                   seed_fields=seed_data)
    df = restore_daily(model.get_synthetic_df(), start, trend_col)
    df.to_csv(f'synthetic-data-{dataset}.csv', index=False)

zredlined

# Create the Gretel Synthtetics Training / Model Configuration

from pathlib import Path
from gretel_helpers.synthetics import create_df, SyntheticDataBundle

config_template = {
    "epochs": 50,
    "early_stopping": False,
    "vocab_size": 20,
    "reset_states": False, 

zredlined

def trends_only(source_df: pd.DataFrame, trend_col: str) -> (float, pd.DataFrame):
    """ Extract trends as training features vs total volume """
    df = source_df.copy()
    start_val = df.at[0, trend_col]
    df.sin = df[[trend_col]].diff()
    df.at[0, trend_col] = 0.00
    return start_val, df


def restore_daily(source_df: pd.DataFrame, start_val: float, trend_col: str):

zredlined

# Load and preview dataset
import datetime
import pandas as pd
import numpy as np

day = 24 * 60 * 60
year = 365.2425 * day


def load_dataframe() -> pd.DataFrame:

zredlined

#!/bin/bash

# Initialize conda environment
echo 'Create Conda env? Type y or n and then press [ENTER]:'
read create_env

if [ $create_env = "y" ];
then
  echo "Provide name for Conda virtualenv and then press [ENTER]:"
  read env_name