Initial commit

.gitignore

@@ -0,0 +1,165 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+*.pyc
+.DS_Store
+build/
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

.python-version

@@ -0,0 +1 @@
+3.11

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 William FH
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,42 @@
+# promptimizer
+
+Prompt optimization trainer.
+
+Example:
+
+```shell
+uv venv
+source .venv/bin/activate
+uv pip install -e .
+python main.py tweet
+```
+
+Script:
+
+```shell
+usage: main.py [-h] [--batch-size BATCH_SIZE] [--train-size TRAIN_SIZE] [--epochs EPOCHS]
+               {scone,tweet,metaprompt,simpleqa}
+
+Optimize prompts for different tasks.
+
+positional arguments:
+  {scone,tweet,metaprompt,simpleqa}
+                        Task to optimize
+
+options:
+  -h, --help            show this help message and exit
+  --batch-size BATCH_SIZE
+                        Batch size for optimization
+  --train-size TRAIN_SIZE
+                        Training size for optimization
+  --epochs EPOCHS       Number of epochs for optimization
+```
+
+Currently has ~ 4 tasks:
+
+- tweet: write tweets
+- simpleqa: really hard Q&A
+- scone: NLI
+- metaprompt: optimize the metaprompt itself over the above 3 tasks
+
+![run](./static/optimizer.gif)

design.md

@@ -0,0 +1,54 @@
+# Design Notes
+
+Something like this:
+```python
+@dataclass
+class Task:
+    """Represents a specific task for prompt optimization."""
+
+    name: str
+    train_datasets: str
+    dev_dataset_name: str
+    test_dataset_name: str
+    evaluators: list[Callable[[Run, Example], dict]] 
+    initial_prompt: str
+    system: Runnable # Prompt + LLM most likely, where the prompt
+    max_iterations: int = 1
+    baseline_experiment: UUID | None
+```
+
+Trainer loop:
+
+1. If baseline_experiment is set, fetch metrics for that, otherwise, run current prompt on dev dataset and get baseline scores. 
+2. Train:
+```python
+- list examples in train dataset
+for x in epochs:
+    - Shuffle examples list & truncate to train size (default None) & split into batches
+    - For batch in batches: 
+        - Run aevaluate on batch
+        - Format feedback
+        - Use current metaprompt to update the task-specific prompt
+```
+3. Test
+    - Run initial/baseline prompt on test dataset and get scores
+    - Run baseline prompt on test dataset and get scores
+    - Print out comparisons
+
+
+
+Then for optimizing the metaprompt, this itself can be framed as a task,but the langsmith dataset would be kinda more of a reference than actual values....
+
+The metaprompt task would be like:
+- train dataset: each example's inputs is the name of a sub-task(?) and subset of the train dataset?
+- system would be some object that:
+initializes with a map of the train dataset to the task objects for the tasks it's trying to optimize on.
+prelim - loads a batch of the subtask-specific examples based on the subset in that example. Looks up evaluators, etc. for this task from the map on the object.
+the evaluator looks at the run outputs
+1. Run initial task prompt over that batch
+2. Run task evals on the results
+3. Run metaprompt to generate new prompt
+4. Run task evals on the new results
+5. return a dict of {original: ..., new: ..., original_scores: ..., new_scores: ....}
+6. eavluator takes those outputs and combines into a singel score. Compare the results and assert they are monitonically improving. And/Or could run a preference evaluator LLM-as-judge.
+    so basically this task's evaluator would be 1 if better, 0.5 if same, 0 if worse (or something; this is just an example)

main.py

@@ -0,0 +1,55 @@
+import asyncio
+import argparse
+
+from prompt_optimizer.trainer import PromptOptimizer
+from langchain_anthropic import ChatAnthropic
+import langsmith as ls
+from prompt_optimizer.tasks.scone import scone_task
+from prompt_optimizer.tasks.tweet_generator import tweet_task
+from prompt_optimizer.tasks.metaprompt import metaprompt_task
+from prompt_optimizer.tasks.simpleqa import simpleqa_task
+
+
+tasks = {
+    "scone": scone_task,
+    "tweet": tweet_task,
+    "metaprompt": metaprompt_task,
+    "simpleqa": simpleqa_task,
+}
+
+
+optimizer = PromptOptimizer(
+    model=ChatAnthropic(model="claude-3-5-sonnet-20241022", max_tokens_to_sample=8192),
+)
+
+
+async def run(task_name: str, batch_size: int, train_size: int, epochs: int):
+    task = tasks.get(task_name)
+    if not task:
+        raise ValueError(f"Unknown task: {task_name}")
+
+    with ls.tracing_context(project_name="Optim"):
+        return await optimizer.optimize_prompt(
+            task, batch_size=batch_size, train_size=train_size, epochs=epochs
+        )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Optimize prompts for different tasks."
+    )
+    parser.add_argument("task", choices=list(tasks), help="Task to optimize")
+    parser.add_argument(
+        "--batch-size", type=int, default=40, help="Batch size for optimization"
+    )
+    parser.add_argument(
+        "--train-size", type=int, default=40, help="Training size for optimization"
+    )
+    parser.add_argument(
+        "--epochs", type=int, default=2, help="Number of epochs for optimization"
+    )
+
+    args = parser.parse_args()
+
+    results = asyncio.run(run(args.task, args.batch_size, args.train_size, args.epochs))
+    print(results)

prompt_optimizer/__init__.py

@@ -0,0 +1,11 @@
+import tqdm
+
+
+# Deactivate tqdm by replacing it with a no-op function
+def tqdm_noop(*args, **kwargs):
+    if args and hasattr(args[0], "__iter__"):
+        return args[0]
+    return args
+
+
+tqdm.auto = tqdm_noop