wzxii's picture
Update index.html
a83107f verified
raw
history blame
11.6 kB
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Memorization or Generation of Big Code Model Leaderboard</title>
<link rel="stylesheet" href="style.css">
<script src="echarts.min.js"></script>
</head>
<body>
<section class="section_title">
<h1>
⭐ <span style="color: rgb(223, 194, 25);">Memorization</span> or
<span style="color: rgb(223, 194, 25);">Generation</span>
of Big
<span style="color: rgb(223, 194, 25);">Code</span>
Models
<span style="color: rgb(223, 194, 25);">Leaderboard</span>
</h1>
<div class="section_title__imgs">
<a href="https://github.com/YihongDong/CDD-TED4LLMs" id="a_github" target="_blank">
<img src="https://img.shields.io/badge/github-%23121011.svg?style=for-the-badge&logo=github&logoColor=white">
</a>
<a href="https://arxiv.org/abs/2402.15938" id="a_arxiv" target="_blank">
<img src="https://img.shields.io/badge/PAPER-ACL'24-ad64d4.svg?style=for-the-badge">
</a>
</div>
<div class="section_title__p">
<p>Inspired from the <a href="https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard" target="_blank">πŸ€— Open LLM Leaderboard</a> and
<a href="https://huggingface.co/spaces/optimum/llm-perf-leaderboard" target="_blank">πŸ€— Open LLM-Perf Leaderboard πŸ‹οΈ</a>,
we compare performance of base code generation models on
<a href="https://huggingface.co/datasets/openai_humaneval" target="_blank">HumanEval</a> and
<a href="https://github.com/YihongDong/CodeGenEvaluation" target="_blank">HumanEval-ET</a> benchamrk. We also measure Memorization-Generalization Index and
provide information about the models.
We compare both open and closed pre-trained code models, that people can start from as base models for
their trainings.
</p>
</div>
</section>
<section class="section_button">
<button id="btn_evalTable">πŸ” Evalution Table</button>
<button id="btn_plot">πŸ“Š Performance Plot</button>
<button id="btn_about">πŸ“ About</button>
<button id="btn_submit">πŸš€ Submit results</button>
</section>
<section class="section_evalTable" id="sec_evalTable">
<div class="section_evalTable__table">
<table id="evalTable">
<colgroup>
<col style="width: 25%">
<col style="width: 15%">
<col style="width: 15%">
<col style="width: 15%">
<col style="width: 15%">
<col style="width: 15%">
</colgroup>
<thead>
<!-- <th rowspan="2">Benchmark</th> -->
<th rowspan="2">Model
<button class="button_sort" data-direction="desc" data-type="name"></button>
</th>
<th data-direction="desc" rowspan="2" data-type="MGI">MGI
<button class="button_sort" data-direction="desc" data-type="MGI"></button>
</th>
<th colspan="2">Pass@1(temp=0)</th>
<th colspan="2">Pass@1(temp=0.8)</th>
<tr>
<th>HumanEval
<button class="button_sort" data-direction="desc" data-type="temp0_HumanEval"></button>
</th>
<th>HumanEval-ET
<button class="button_sort" data-direction="desc" data-type="temp0_HumanEval_ET"></button>
</th>
<th>HumanEval
<button class="button_sort" data-direction="desc" data-type="temp0_8_HumanEval"></button>
</th>
<th>HumanEval-ET
<button class="button_sort" data-direction="desc" data-type="temp0_8_HumanEval_ET"></button>
</th>
</tr>
</thead>
<tbody>
</tbody>
</table>
</table>
<script src="table.js"></script>
</div>
<div class="section_evalTable__notes">
<p><strong>Notes</strong>
<p>
<ul>
<li>MGI stands for Memorization-Generalization Index, which is derived from Avg. Peak in the original paper. A higher MGI value indicates a greater propensity for a model to engage in memorization as opposed to generalization.</li>
<li>For more details check the πŸ“ About section.</li>
</ul>
</div>
</section>
<section class="section_plot" id="sec_plot">
<div style="display: flex;">
<div class="section_plot__div" id="sec_plot__div1">
<div class="section_plot__btnGroup" id="sec_plot__btnGroup1">
<button id="btn_temp0_HumanEval"></button>
<span id="span_temp0_HumanEval">Pass@1 (temp = 0)</span>
<button id="btn_temp0_HumanEval_ET"></button>
<span id="span_temp0_HumanEval_ET">Pass@1 (temp = 0.8)</span>
</div>
<div id="sec_plot__chart1" style="width:706.5px; height:550px;"></div>
</div>
<div class="section_plot__div" id="sec_plot__div2">
<div class="section_plot__btnGroup" id="sec_plot__btnGroup2">
<button id="btn_temp0_8_HumanEval"></button>
<span id="span_temp0_8_HumanEval">Pass@1 (temp = 0)</span>
<button id="btn_temp0_8_HumanEval_ET"></button>
<span id="span_temp0_8_HumanEval_ET">Pass@1 (temp = 0.8)</span>
</div>
<div id="sec_plot__chart2" style="width:706.5px; height:550px;"></div>
</div>
</div>
<script src="chart.js"></script>
</section>
<section class="section_about" id="sec_about">
<h2>Context</h2>
<div>
<p>The growing number of code models released by the community necessitates a comprehensive evaluation to
reliably benchmark their capabilities.
Similar to the πŸ€— Open LLM Leaderboard, we selected two common benchmarks for evaluating Code LLMs on
multiple programming languages:</p>
<ul>
<li>HumanEval - benchmark for measuring functional correctness for synthesizing programs from
docstrings. It consists of 164 Python programming problems.</li>
<li>MultiPL-E - Translation of HumanEval to 18 programming languages.</li>
<li>Throughput Measurement - In addition to these benchmarks, we also measure model throughput on a
batch size of 1 and 50 to compare their inference speed.</li>
</ul>
<h3>Benchmark & Prompts</h3>
<ul>
<li>HumanEval-Python reports the pass@1 on HumanEval; the rest is from MultiPL-E benchmark.</li>
<li>For all languages, we use the original benchamrk prompts for all models except HumanEval-Python,
where we separate base from instruction models.
We use the original code completion prompts for HumanEval for all base models, but for Instruction
models,
we use the Instruction version of HumanEval in HumanEvalSynthesize delimited by the tokens/text
recommended by the authors of each model
(we also use a max generation length of 2048 instead of 512).</li>
</ul>
<p>Figure below shows the example of OctoCoder vs Base HumanEval prompt, you can find the other prompts
here.</p>
</div>
<div>
<p>- An exception to this is the Phind models. They seem to follow to base prompts better than the
instruction versions.
Therefore, following the authors' recommendation we use base HumanEval prompts without stripping them of
the last newline.
- Also note that for WizardCoder-Python-34B-V1.0 & WizardCoder-Python-13B-V1.0 (CodeLLaMa based),
we use the HumanEval-Python instruction prompt that the original authors used with their postprocessing
(instead of HumanEvalSynthesize),
code is available [here](https://github.com/bigcode-project/bigcode-evaluation-harness/pull/133).</p>
<h3>Evalution Parameters</h3>
<ul>
<li>All models were evaluated with the bigcode-evaluation-harness with top-p=0.95, temperature=0.2,
max_length_generation 512, and n_samples=50.</li>
</ul>
<h3>Throughput and Memory Usage</h3>
<ul>
<li>Throughputs and peak memory usage are measured using Optimum-Benchmark which powers Open LLM-Perf
Leaderboard. (0 throughput corresponds to OOM).</li>
</ul>
<h3>Scoring and Rankings</h3>
<ul>
<li>Average score is the average pass@1 over all languages. For Win Rate, we find model rank for each
language and compute num_models - (rank -1), then average this result over all languages.</li>
</ul>
<h3>Miscellaneous</h3>
<ul>
<li>#Languages column represents the number of programming languages included during the pretraining.
UNK means the number of languages is unknown.</li>
</ul>
</div>
</section>
<section class="section_submit" id="sec_submit">
<h2>How to submit models/results to the leaderboard?</h2>
<div>
<p>We welcome the community to submit evaluation results of new models. These results will be added as
non-verified, the authors are however required to upload their generations in case other members want to
check.</p>
<h3>1 - Running Evaluation</h3>
<p>We wrote a detailed guide for running the evaluation on your model. You can find the it in
bigcode-evaluation-harness/leaderboard. This will generate a json file summarizing the results, in
addition to the raw generations and metric files.</p>
<h3>2- Submitting Results πŸš€</h3>
<p>To submit your results create a Pull Request in the community tab to add them under the folder
community_results in this repository:</p>
<ul>
<li>Create a folder called ORG_MODELNAME_USERNAME for example bigcode_starcoder_loubnabnl</li>
<li>Put your json file with grouped scores from the guide, in addition generations folder and metrics
folder in it.</li>
</ul>
<p>The title of the PR should be [Community Submission] Model: org/model, Username: your_username, replace
org and model with those corresponding to the model you evaluated.</p>
</div>
</section>
<footer>
</footer>
<script src="button.js"></script>
</body>
</html>