| { |
| "sha": "3025a531af7450e7df8b96fe0440e9804480bbad", |
| "node_id": "C_kwDOHttWl9oAKDMwMjVhNTMxYWY3NDUwZTdkZjhiOTZmZTA0NDBlOTgwNDQ4MGJiYWQ", |
| "commit": { |
| "author": { |
| "name": "Rudolf Adamkovič", |
| "email": "rudolf@adamkovic.org", |
| "date": "2026-01-23T13:35:15Z" |
| }, |
| "committer": { |
| "name": "Arjun Guha", |
| "email": "arjunguha@users.noreply.github.com", |
| "date": "2026-01-28T13:36:52Z" |
| }, |
| "message": "Clean up whitespace in README", |
| "tree": { |
| "sha": "5654a4b922d2d0099553f4ce01fccde3b2963df5", |
| "url": "https://api.github.com/repos/nuprl/MultiPL-E/git/trees/5654a4b922d2d0099553f4ce01fccde3b2963df5" |
| }, |
| "url": "https://api.github.com/repos/nuprl/MultiPL-E/git/commits/3025a531af7450e7df8b96fe0440e9804480bbad", |
| "comment_count": 0, |
| "verification": { |
| "verified": false, |
| "reason": "unsigned", |
| "signature": null, |
| "payload": null, |
| "verified_at": null |
| } |
| }, |
| "url": "https://api.github.com/repos/nuprl/MultiPL-E/commits/3025a531af7450e7df8b96fe0440e9804480bbad", |
| "html_url": "https://github.com/nuprl/MultiPL-E/commit/3025a531af7450e7df8b96fe0440e9804480bbad", |
| "comments_url": "https://api.github.com/repos/nuprl/MultiPL-E/commits/3025a531af7450e7df8b96fe0440e9804480bbad/comments", |
| "author": { |
| "login": "rudolf-adamkovic", |
| "id": 1158274, |
| "node_id": "MDQ6VXNlcjExNTgyNzQ=", |
| "avatar_url": "https://avatars.githubusercontent.com/u/1158274?v=4", |
| "gravatar_id": "", |
| "url": "https://api.github.com/users/rudolf-adamkovic", |
| "html_url": "https://github.com/rudolf-adamkovic", |
| "followers_url": "https://api.github.com/users/rudolf-adamkovic/followers", |
| "following_url": "https://api.github.com/users/rudolf-adamkovic/following{/other_user}", |
| "gists_url": "https://api.github.com/users/rudolf-adamkovic/gists{/gist_id}", |
| "starred_url": "https://api.github.com/users/rudolf-adamkovic/starred{/owner}{/repo}", |
| "subscriptions_url": "https://api.github.com/users/rudolf-adamkovic/subscriptions", |
| "organizations_url": "https://api.github.com/users/rudolf-adamkovic/orgs", |
| "repos_url": "https://api.github.com/users/rudolf-adamkovic/repos", |
| "events_url": "https://api.github.com/users/rudolf-adamkovic/events{/privacy}", |
| "received_events_url": "https://api.github.com/users/rudolf-adamkovic/received_events", |
| "type": "User", |
| "user_view_type": "public", |
| "site_admin": false |
| }, |
| "committer": { |
| "login": "arjunguha", |
| "id": 20065, |
| "node_id": "MDQ6VXNlcjIwMDY1", |
| "avatar_url": "https://avatars.githubusercontent.com/u/20065?v=4", |
| "gravatar_id": "", |
| "url": "https://api.github.com/users/arjunguha", |
| "html_url": "https://github.com/arjunguha", |
| "followers_url": "https://api.github.com/users/arjunguha/followers", |
| "following_url": "https://api.github.com/users/arjunguha/following{/other_user}", |
| "gists_url": "https://api.github.com/users/arjunguha/gists{/gist_id}", |
| "starred_url": "https://api.github.com/users/arjunguha/starred{/owner}{/repo}", |
| "subscriptions_url": "https://api.github.com/users/arjunguha/subscriptions", |
| "organizations_url": "https://api.github.com/users/arjunguha/orgs", |
| "repos_url": "https://api.github.com/users/arjunguha/repos", |
| "events_url": "https://api.github.com/users/arjunguha/events{/privacy}", |
| "received_events_url": "https://api.github.com/users/arjunguha/received_events", |
| "type": "User", |
| "user_view_type": "public", |
| "site_admin": false |
| }, |
| "parents": [ |
| { |
| "sha": "a8081e1fce899f00f157fdc6bf8e47677438f7ff", |
| "url": "https://api.github.com/repos/nuprl/MultiPL-E/commits/a8081e1fce899f00f157fdc6bf8e47677438f7ff", |
| "html_url": "https://github.com/nuprl/MultiPL-E/commit/a8081e1fce899f00f157fdc6bf8e47677438f7ff" |
| } |
| ], |
| "stats": { |
| "total": 116, |
| "additions": 58, |
| "deletions": 58 |
| }, |
| "files": [ |
| { |
| "sha": "6781e1fac78763ddb3c7c21486bb2be838c8cac6", |
| "filename": "README.md", |
| "status": "modified", |
| "additions": 58, |
| "deletions": 58, |
| "changes": 116, |
| "blob_url": "https://github.com/nuprl/MultiPL-E/blob/3025a531af7450e7df8b96fe0440e9804480bbad/README.md", |
| "raw_url": "https://github.com/nuprl/MultiPL-E/raw/3025a531af7450e7df8b96fe0440e9804480bbad/README.md", |
| "contents_url": "https://api.github.com/repos/nuprl/MultiPL-E/contents/README.md?ref=3025a531af7450e7df8b96fe0440e9804480bbad", |
| "patch": "@@ -21,13 +21,13 @@ For more information:\n \n ## Tutorial\n \n-These are instructions on how to use MultiPl-E directly, without the \n+These are instructions on how to use MultiPl-E directly, without the\n BigCode evaluation harness.\n \n In this tutorial, we will run a small experiment to evaluate the performance of\n-[SantaCoder] on Rust with a small subset of the MBPP benchmarks. \n+[SantaCoder] on Rust with a small subset of the MBPP benchmarks.\n We will only fetch 20 completions per problem, so that you\n-can run it quickly on a single machine. \n+can run it quickly on a single machine.\n You can also run on the full suite of benchmarks or substitute your own\n benchmark programs. Later, we'll show you how to add support for other languages\n and evaluate other models.\n@@ -44,7 +44,7 @@ and evaluate other models.\n \n 3. You need to install one of [Podman] or [Docker].\n \n-3. Check out the repository: \n+3. Check out the repository:\n \n ```bash\n git clone https://github.com/nuprl/MultiPL-E\n@@ -58,7 +58,7 @@ and evaluate other models.\n \n ### Background\n \n-Out of the box, MultiPL-E supports several models, programming languages, \n+Out of the box, MultiPL-E supports several models, programming languages,\n and datasets. Using MultiPL-E is a two step process:\n \n 1. We *generate* completions, which requires a GPU.\n@@ -92,7 +92,7 @@ Hugging Face Hub. You can use any other text generation model instead.\n \n Notes:\n \n-1. This command requires about 13 GB VRAM and takes 30 minutes with a Quadro RTX \n+1. This command requires about 13 GB VRAM and takes 30 minutes with a Quadro RTX\n 6000.\n 2. If you have less VRAM, you can set `--batch-size` to a smaller value.\n E.g., with `--batch-size 10` it should work on consumer graphics cards,\n@@ -168,7 +168,7 @@ can regenerate with `--temperature 0.8`.\n \n **Warning:** In generation, we used `--completion-limit 20` to only generate\n 20 samples for each prompt. You should remove this flag to generate 200 samples\n-for temperature 0.8. We have found that 20 samples is adequate for estimate \n+for temperature 0.8. We have found that 20 samples is adequate for estimate\n pass@1 (there will be a little variance). However, you need more samples to estimate\n pass@10 and pass@100.\n \n@@ -194,41 +194,41 @@ In general, you need to make three changes to support a new language *L*:\n \n ### Writing the Translator\n \n-Let's say we had not included Perl in the set of benchmark languages and \n-you want to add it. In a new file `humaneval_to_perl.py` you will need to \n+Let's say we had not included Perl in the set of benchmark languages and\n+you want to add it. In a new file `humaneval_to_perl.py` you will need to\n define a class called `Translator`. `Translator` contains numerous methods -\n-the interface for a generic `Translator` class is provided in `base_language_translator.py `. \n+the interface for a generic `Translator` class is provided in `base_language_translator.py `.\n \n *Note*: You must name your translator `humaneval_to_L.py`. However, the code\n works with several other benchmarks, including MBPP.\n \n-There are three types of methods for `Translator`: (1) methods that handle \n+There are three types of methods for `Translator`: (1) methods that handle\n translating the prompt, (2) methods that handle translating the unit tests, and\n-(3) methods that handle the value-to-value translation. \n+(3) methods that handle the value-to-value translation.\n \n-First, let's handle converting the Python prompt to a Perl prompt. This is \n-done by the `translate_prompt` method. `translate_prompt` needs to return \n-a string (we definitely suggest using a formatted Python string here) that \n-contains the Perl prompt and then the Perl function signature. We suggest \n-accumulating the prompt into one string as follows: \n+First, let's handle converting the Python prompt to a Perl prompt. This is\n+done by the `translate_prompt` method. `translate_prompt` needs to return\n+a string (we definitely suggest using a formatted Python string here) that\n+contains the Perl prompt and then the Perl function signature. We suggest\n+accumulating the prompt into one string as follows:\n ```\n perl_description = \"# \" + re.sub(DOCSTRING_LINESTART_RE, \"\\n# \", description.strip()) + \"\\n\"\n ```\n-where `\"#\"` are Perl single-line comments. `DOCSTRING_LINESTART_RE` identifies the \n-first line in the prompt using a regex and then `description` is a string representing \n-the rest of the prompt. This process should be pretty simple - just connect them together with \n+where `\"#\"` are Perl single-line comments. `DOCSTRING_LINESTART_RE` identifies the\n+first line in the prompt using a regex and then `description` is a string representing\n+the rest of the prompt. This process should be pretty simple - just connect them together with\n your comment structure of choice.\n \n-The argument `name` to `translate_prompt` takes care of the function name, you \n-just need to format the function arguments (argument `args`) and delimiters to complete \n+The argument `name` to `translate_prompt` takes care of the function name, you\n+just need to format the function arguments (argument `args`) and delimiters to complete\n the prompt translation.\n \n Now let's consider the three methods which help translate unit tests:\n-`test_suite_prefix_lines`, `test_suite_suffix_lines`, and `deep_equality`. \n-The prefix and suffix methods return a \"wrapper\" around the set of generated unit \n-tests. In most languages, as is the case in Perl, the prefix defines a function/class \n-for testing and the suffix calls that function. This may include calls to your testing library \n-of choice (please look at existing `humaneval_to` files for examples!). \n+`test_suite_prefix_lines`, `test_suite_suffix_lines`, and `deep_equality`.\n+The prefix and suffix methods return a \"wrapper\" around the set of generated unit\n+tests. In most languages, as is the case in Perl, the prefix defines a function/class\n+for testing and the suffix calls that function. This may include calls to your testing library\n+of choice (please look at existing `humaneval_to` files for examples!).\n The wrapper in Perl we use is:\n ```\n sub testhumaneval {\n@@ -238,19 +238,19 @@ sub testhumaneval {\n testhumaneval();\n ```\n \n-Note the argument `entry_point` to `test_suite_prefix_lines`: this is the name \n-of the function for each benchmark. In most languages, we either assign that to \n-a variable `candidate` (as done in the original HumanEval benchmark) or call \n-`entry_point` directly. \n+Note the argument `entry_point` to `test_suite_prefix_lines`: this is the name\n+of the function for each benchmark. In most languages, we either assign that to\n+a variable `candidate` (as done in the original HumanEval benchmark) or call\n+`entry_point` directly.\n \n-The final unit test function is `deep_equality`, which is where you define how \n+The final unit test function is `deep_equality`, which is where you define how\n to check whether two arguments (`left` and `right`) are structurally equal. In Perl\n-we do this with `eq_deeply`. (Hint: note that sometimes the order of `left` and \n-`right` can be switched in some testing frameworks - try this out to produce \n+we do this with `eq_deeply`. (Hint: note that sometimes the order of `left` and\n+`right` can be switched in some testing frameworks - try this out to produce\n the best error messages possible!).\n \n Third, let's tackle the value-to-value translation methods. All of them take\n-a Python value (or some representation of one) as an argument and return a string \n+a Python value (or some representation of one) as an argument and return a string\n representing that value's equivalent in Perl.\n \n For instance, `gen_dict` defines what dictionaries in Python should map to in\n@@ -262,35 +262,35 @@ nstead of `:` to differentiate keys and values in Perl.\n return \"{\" + \", \".join(f\"{k} => {v}\" for k, v in zip(keys, values)) + \"}\"\n ```\n \n-This step should be quite straightforward for each value and its associated \n-method. When there is choice, we used our language knowledge or consulted \n-the style guides from the language communities (see our paper's Appendix). As we \n-mention in our paper, the ease of value-to-value mapping is one of the key aspects of \n-this approach. \n+This step should be quite straightforward for each value and its associated\n+method. When there is choice, we used our language knowledge or consulted\n+the style guides from the language communities (see our paper's Appendix). As we\n+mention in our paper, the ease of value-to-value mapping is one of the key aspects of\n+this approach.\n \n There are also smaller elements to `Translator` (stop tokens, file_ext, etc.)\n-that you will need to populate accordingly. \n+that you will need to populate accordingly.\n \n-If you've successfully gotten to this point: great, you're done and can move \n-on to `eval_foo` and testing. If you wanted to add a statically typed \n+If you've successfully gotten to this point: great, you're done and can move\n+on to `eval_foo` and testing. If you wanted to add a statically typed\n benchmark - Read on!\n \n #### What about statically typed languages?\n \n-Statically typed translations are notably more challenging to implement than the \n-Perl example above. Rather than walk you through the steps directly, we provide a \n+Statically typed translations are notably more challenging to implement than the\n+Perl example above. Rather than walk you through the steps directly, we provide a\n well-documented version of `humaneval_to_ts.py` for TypeScript as an example. Feel free\n-to also consult translations for other languages in the benchmark, although your \n-mileage may vary. \n+to also consult translations for other languages in the benchmark, although your\n+mileage may vary.\n \n ### Writing the Execution Script\n \n-Now that you're done converting Python to your language of choice, you need \n-to define how to evaluate the generated programs. As a reminder, one of the \n+Now that you're done converting Python to your language of choice, you need\n+to define how to evaluate the generated programs. As a reminder, one of the\n contributions of this benchmark suite is actually evaluating the generated\n code. Let's continue with the idea that you are adding Perl as a new language to our dataset.\n \n-In `eval_L.py` you should define a function, `eval_script`, with the \n+In `eval_L.py` you should define a function, `eval_script`, with the\n following signature and imports:\n ```\n from pathlib import Path\n@@ -299,19 +299,19 @@ from safe_subprocess import run\n def eval_script(path: Path):\n ```\n \n-In the body of `eval_script` you should call `run` with the \n+In the body of `eval_script` you should call `run` with the\n requisite arguments (please refer to it's documentation and your computing architecture\n to do this correctly). For our results, we use the following call to `run` for Perl:\n ```\n r = run([\"perl\", path])\n ```\n \n-You should then determine how to handle what gets assigned to `r`. If you \n+You should then determine how to handle what gets assigned to `r`. If you\n look around the eval scripts we provide, there are different granularities for\n handling program evaluation. For instance some statically typed errors\n handle compilation and runtime errors differently. We recommend, at minimum,\n-handling success (typically exit code 0), timeouts, syntax errors, \n-and exceptions as four subclasses of results. You can do this using \n+handling success (typically exit code 0), timeouts, syntax errors,\n+and exceptions as four subclasses of results. You can do this using\n `try-except` statements or simply with conditionals:\n \n ```\n@@ -322,7 +322,7 @@ and exceptions as four subclasses of results. You can do this using\n status = \"OK\"\n ```\n \n-`eval_script` should return a dictionary of the form below - the scripts above \n+`eval_script` should return a dictionary of the form below - the scripts above\n rely on this output format to calculate pass@k metrics:\n \n ```\n@@ -341,7 +341,7 @@ The final two steps are:\n 2. Create a Dockerfile for your language in the `evaluation` directory.\n \n There is one final step if you want to run the completion\n-tutorial above for your brand new language. Open `containerized_eval.py` and \n+tutorial above for your brand new language. Open `containerized_eval.py` and\n add links to your new language in two places:\n \n ### Writing the Terms to Translate Comments\n@@ -417,7 +417,7 @@ def my_function(a: int, b: int, c: int, k: int) -> int:\n (a ** n) + (b ** n) = (c ** n).\n \"\"\"\n pass\n- \n+\n \n ### Unit tests below ###\n def check(candidate):\n@@ -466,7 +466,7 @@ python3 prepare_prompts_for_hfhub.py \\\n --output ../L_prompts.jsonl\n ```\n \n-You can then test the dataset by following the steps in \n+You can then test the dataset by following the steps in\n [Testing a new language](https://github.com/nuprl/MultiPL-E?tab=readme-ov-file#testing-a-new-language).\n \n ## Updating MultiPL-E" |
| } |
| ] |
| } |
|
|
