chore: upload paper reference impl

reference/README.md

@@ -0,0 +1,7 @@
+# Reference Implementations
+
+This directory contains reference implementations of experiments run in the FanOutQA paper. These implementations are
+provided "as-is", with no guarantees on usability or support provided. Note the code quality of these implementations
+may be lower than the library code, as these are earlier explorations run a) on a deadline and b) to develop the 
+interfaces that were eventually exposed in this library.
+

reference/paper_benchmarks/.gitignore

@@ -0,0 +1,161 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# 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
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__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 maintainted 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/
+
+**.DS_Store
+results/extra-*
+results/prompts-*
+results-old/
+.wikicache/
+tmp*/
+BLEURT-20/
+.llmcache/

reference/paper_benchmarks/README.md

@@ -0,0 +1,28 @@
+# benchmarks
+
+make sure to do `python -m spacy download en_core_web_sm`
+
+closedbook: just the q and 0 shot prompt to output short text
+
+openbook: like closedbook but you can search wikipedia wowza
+
+wiki provided: chunk the known evidence 1024 characters at a time (preferring splitting at paragraph and sentence
+boundaries), retrieve as many passages fit into context window as possible, use bm25
+
+run:
+
+```shell
+WIKI_CACHE_DIR=/nlp/data/andrz/fanoutqa-bench/.wikicache python run_openbook.py -m mistral-chat -n 3
+for i in slurm/*; do sbatch $i; done
+```
+
+# eval
+
+install:
+
+```shell
+wget https://storage.googleapis.com/bleurt-oss-21/BLEURT-20.zip
+unzip BLEURT-20.zip
+rm BLEURT-20.zip
+python -m spacy download en_core_web_sm
+```

reference/paper_benchmarks/bench/engines.py

@@ -0,0 +1,111 @@
+import os
+
+import aiolimiter
+import torch
+from kani import AIFunction, ChatMessage
+from kani.engines.anthropic import AnthropicEngine
+from kani.engines.base import BaseEngine
+from kani.engines.huggingface.llama2 import LlamaEngine
+from kani.engines.openai import OpenAIEngine
+from kani.engines.openai.models import ChatCompletion
+
+# openai
+ft_org_id = "org-dxkPAvuWroRmYw7z0VhqMC5S"
+lab_org_id = "org-bgAXfs8WdU5942SLngg0OGpd"
+
+# hf
+common_hf_model_args = dict(
+    device_map="auto",
+    torch_dtype=torch.float16,
+)
+
+if mcd := os.getenv("MODEL_CACHE_DIR"):
+    common_hf_model_args["cache_dir"] = mcd
+
+
+def can_parallel(engine):
+    return isinstance(engine, OpenAIEngine)
+
+
+def get_engine(name: str, ctx_size=None) -> BaseEngine:
+    if name == "gpt-4":
+        return RatelimitedOpenAIEngine(
+            model="gpt-4-0613",
+            organization=lab_org_id,
+            max_context_size=ctx_size or 8192,
+            tpm_limiter=aiolimiter.AsyncLimiter(300000),
+            temperature=0,
+        )
+    elif name == "gpt-4-turbo":
+        return RatelimitedOpenAIEngine(
+            model="gpt-4-0125-preview",
+            organization=ft_org_id,
+            max_context_size=ctx_size or 128000,
+            tpm_limiter=aiolimiter.AsyncLimiter(400000),
+            temperature=0,
+        )
+    elif name == "gpt-3.5-turbo":
+        return RatelimitedOpenAIEngine(
+            model="gpt-3.5-turbo-1106",
+            organization=ft_org_id,
+            max_context_size=ctx_size or 16000,
+            tpm_limiter=aiolimiter.AsyncLimiter(2000000),
+            temperature=0,
+        )
+    elif name == "llama-chat":
+        return LlamaEngine(
+            "meta-llama/Llama-2-70b-chat-hf",
+            max_context_size=ctx_size or 4096,
+            model_load_kwargs=common_hf_model_args,
+            use_auth_token=True,
+            do_sample=False,
+        )
+    elif name == "mistral-chat":
+        return LlamaEngine(
+            "mistralai/Mistral-7B-Instruct-v0.2",
+            max_context_size=ctx_size or 32768,
+            model_load_kwargs=common_hf_model_args,
+            do_sample=False,
+        )
+    elif name == "mixtral":
+        return LlamaEngine(
+            "mistralai/Mixtral-8x7B-Instruct-v0.1",
+            max_context_size=ctx_size or 32768,
+            model_load_kwargs=common_hf_model_args,
+            do_sample=False,
+        )
+    elif name == "longllama":
+        return LlamaEngine(
+            "syzymon/long_llama_3b_instruct",
+            max_context_size=ctx_size or 512000,
+            tokenizer_kwargs={"trust_remote_code": True},
+            model_load_kwargs={"trust_remote_code": True, **common_hf_model_args},
+            do_sample=False,
+        )
+    elif name == "claude":
+        return AnthropicEngine(
+            model="claude-2.1",
+            max_context_size=ctx_size or 200000,
+            temperature=0,
+        )
+    else:
+        raise ValueError("Invalid model name")
+
+
+class RatelimitedOpenAIEngine(OpenAIEngine):
+    def __init__(
+        self, *args, rpm_limiter: aiolimiter.AsyncLimiter = None, tpm_limiter: aiolimiter.AsyncLimiter = None, **kwargs
+    ):
+        super().__init__(*args, **kwargs)
+        self.rpm_limiter = rpm_limiter
+        self.tpm_limiter = tpm_limiter
+
+    async def predict(
+        self, messages: list[ChatMessage], functions: list[AIFunction] | None = None, **hyperparams
+    ) -> ChatCompletion:
+        if self.rpm_limiter:
+            await self.rpm_limiter.acquire()
+        if self.tpm_limiter:
+            n_toks = self.function_token_reserve(functions) + sum(self.message_len(m) for m in messages)
+            await self.tpm_limiter.acquire(n_toks)
+        return await super().predict(messages, functions, **hyperparams)