Add initial mc-answer-prompt-experiment features

lm_eval/base.py

@@ -513,43 +513,8 @@ class Task(abc.ABC):
         example = self.doc_to_text(doc)
         return description + labeled_examples + example
 
-
-class MultipleChoiceTask(Task, abc.ABC):
-    def doc_to_target(self, doc):
-        return " " + doc['choices'][doc['gold']]
-
-    def construct_requests(self, doc, ctx):
-        lls = [
-            rf.loglikelihood(ctx, " {}".format(choice))[0]
-            for choice in doc['choices']
-        ]
-
-        return lls
-
-    def process_results(self, doc, results):
-        gold = doc["gold"]
-
-        acc = 1. if np.argmax(results) == gold else 0.
-        completion_len = np.array([float(len(i)) for i in doc["choices"]])
-        acc_norm = 1. if np.argmax(results / completion_len) == gold else 0.
-
-        return {
-            "acc": acc,
-            "acc_norm": acc_norm,
-        }
-    
-    def higher_is_better(self):
-        return {
-            "acc": True,
-            "acc_norm": True,
-        }
-    
-    def aggregation(self):
-        return {
-            "acc": mean,
-            "acc_norm": mean,
-        }
-
+from lm_eval.mctask_experimental import MULTIPLE_CHOICE_TASK
+MultipleChoiceTask = MULTIPLE_CHOICE_TASK
 
 class PerplexityTask(Task, abc.ABC):
 

lm_eval/evaluator.py

 import collections
 import itertools
+import os
 import random
 import lm_eval.metrics
 import lm_eval.models
@@ -107,6 +108,9 @@ def evaluate(lm, task_dict, provide_description=None, num_fewshot=0, limit=None,
         Dictionary of results
     """
     # TODO: completely refactor this entire function to not be a huge mess, ideally breaking it down into smaller pieces
+    print(f"{'='*20}")
+    print(f"Task Module: {lm_eval.base.MultipleChoiceTask.__name__}")
+    print(f"{'='*20}")
 
     # TODO: todo: implement proper description-providing system
     assert not provide_description  # not implemented.

lm_eval/mctask_experimental.py

+""" Multiple Choice Format Experiments.
+TODO: Generalize the formatting of fewshot examples.
+"""
+import os
+import abc
+import hashlib
+from argparse import ArgumentError
+from dataclasses import dataclass
+import typing
+from attr import field
+import numpy as np
+import lm_eval.base as base
+from lm_eval.metrics import mean
+
+@dataclass
+class MultipleChoiceDoc:
+    question: str
+    # The possible answer keys, e.g. `["A", "B", "C", "D"]`.
+    # These should be the type as gold?
+    keys: typing.List[str]
+    options: typing.List[str]
+    gold: int
+    id: int = field(init=False)
+
+    def __post_init__(self):
+        self.id = hashlib.sha224(self.question.encode('utf-8')).hexdigest()
+
+class BaseMultipleChoiceTask(base.Task, abc.ABC):
+
+    def doc_to_text(self, doc: MultipleChoiceDoc):
+        return self.format_prompt(doc)
+
+    @abc.abstractclassmethod
+    def format_prompt(cls, doc: MultipleChoiceDoc) -> str:
+        pass
+
+    @abc.abstractmethod
+    def doc_to_target(self, doc: MultipleChoiceDoc) -> str:
+        pass
+
+    @abc.abstractmethod
+    def loglikelihood_continuation(self, doc: MultipleChoiceDoc) -> typing.List[str]:
+        pass
+
+    def construct_requests(self, doc: MultipleChoiceDoc, ctx: str):
+        lls = []
+        conts = self.loglikelihood_continuation(doc)
+        for cont in conts:
+            lls.append(base.rf.loglikelihood(ctx, f" {cont}")[0])
+        return lls
+
+    def process_results(self, doc: MultipleChoiceDoc, results: typing.List):
+        gold = doc.gold
+        ans = np.argmax(results)
+        is_correct = 1. if ans == gold else 0.
+        # Normalize by completion length.
+        conts = self.loglikelihood_continuation(doc)
+        completion_len = np.array([float(len(i)) for i in conts])
+        acc_norm = 1. if np.argmax(results / completion_len) == gold else 0.
+        return {
+            "acc": is_correct,
+            "acc_norm": acc_norm,
+            # Bundle answers: (id, answer key, answer index, is correct).
+            "answer_bundle": (doc.id, doc.keys[ans], ans, is_correct),
+        }
+
+    def higher_is_better(self):
+        return {
+            "acc": True,
+            "acc_norm": True,
+            "answer_bundle": True,
+        }
+
+    def aggregation(self):
+        return {
+            "acc": mean,
+            "acc_norm": mean,
+            "answer_bundle": answer_bundle
+        }
+
+def answer_bundle(items):
+    """ Bundles answers into a csv file. """
+    from pathlib import Path
+    import csv
+    cols = ["question_id", "model_answer", "model_answer_index", "is_correct"]
+    rows = [*items]
+    path = os.environ["QUESTION_RESULT_PATH"]
+    with open(f'{path}/question-by-question-results.csv', 'w') as f:
+        write = csv.writer(f)
+        write.writerow(cols)
+        write.writerows(rows)
+    return 0
+
+
+def key2num(doc: MultipleChoiceDoc, key: str) -> int:
+    return str(doc.keys.index(key) + 1)  # `+ 1` for 1-based indexing.
+
+
+def format_key(key: str, type: str):
+    """ Formats a multiple choice key. E.g.
+        format_key("A", "period") => "A."
+        format_key("A", "parens") => "(A)"
+        format_key("A",  "colon") => "A:"
+    Args:
+    - type: "period" | "parens" | "colon"
+    """
+    if type == "parens":
+        return f"({key})"
+    elif type == "period":
+        return f"{key}."
+    elif type == "colon":
+        return f"{key}:"
+    else:
+        raise ArgumentError()
+
+
+class MC_NoOptionList_OptionLL_Task(BaseMultipleChoiceTask):
+    """
+    Format:
+        Question: <question>
+        Answer: 
+    Continuation:
+        loglikelihood_continuation = <option_i>
+    """
+    def format_prompt(cls, doc: MultipleChoiceDoc) -> str:
+        prompt = "Question: " + doc.question + "\n"
+        prompt += "Answer:"
+        return prompt
+       # return _format_prompt(doc, list_options=False)
+
+    def doc_to_target(self, doc: MultipleChoiceDoc) -> str:
+        return " " + doc.options[doc.gold]
+
+    def loglikelihood_continuation(self, doc: MultipleChoiceDoc) -> typing.List[str]:
+        return [option for option in doc.options]
+
+
+class MC_WithOptionList_OptionLL_Task(BaseMultipleChoiceTask):
+    """
+    Format:
+        Question: <question>
+        <key1>: <option1>
+        <key2>: <option2>
+        ...
+        Answer: 
+    Continuation:
+        loglikelihood_continuation = <option_i>
+    """
+    def format_prompt(cls, doc: MultipleChoiceDoc) -> str:
+        prompt = "Question: " + doc.question + "\n"
+        prompt += "\n".join([
+            f"{format_key(doc.keys[i], 'colon')} {option}"
+            for i, option in enumerate(doc.options)
+        ])
+        prompt += "\nAnswer:"
+        return prompt
+
+    def doc_to_target(self, doc: MultipleChoiceDoc) -> str:
+        return " " + doc.options[doc.gold]
+
+    def loglikelihood_continuation(self, doc: MultipleChoiceDoc) -> typing.List[str]:
+        return [option for option in doc.options]
+
+
+class MC_WithOptionList_LetterLL_Task(BaseMultipleChoiceTask):
+    """
+    Format:
+        Question: <question>
+        <key1>: <option1>
+        <key2>: <option2>
+        ...
+        Answer: 
+    Continuation:
+        loglikelihood_continuation = <key_i>
+    """
+    def format_prompt(cls, doc: MultipleChoiceDoc) -> str:
+        prompt = "Question: " + doc.question + "\n"
+        prompt += "\n".join([
+            f"{format_key(doc.keys[i], 'colon')} {option}"
+            for i, option in enumerate(doc.options)
+        ])
+        prompt += "\nAnswer:"
+        return prompt
+
+    def doc_to_target(self, doc: MultipleChoiceDoc) -> str:
+        return " " + doc.keys[doc.gold]
+
+    def loglikelihood_continuation(self, doc: MultipleChoiceDoc) -> typing.List[str]:
+        return [key for key in doc.keys]
+
+
+class MC_WithOptionList_NumLL_Task(BaseMultipleChoiceTask):
+    """
+    Format:
+        Question: <question>
+        1: <option1>
+        2: <option2>
+        ...
+        Answer: 
+    Continuation:
+        loglikelihood_continuation = <key2num(key_i)>
+    """
+    def format_prompt(cls, doc: MultipleChoiceDoc) -> str:
+        prompt = "Question: " + doc.question + "\n"
+        prompt += "\n".join([
+            f"{format_key(key2num(doc, doc.keys[i]), 'colon')} {option}"
+            for i, option in enumerate(doc.options)
+        ])
+        prompt += "\nAnswer:"
+        return prompt
+
+    def doc_to_target(self, doc: MultipleChoiceDoc) -> str:
+        return f" {doc.gold + 1}"  # `+ 1` for 1-based indexing.
+
+    def loglikelihood_continuation(self, doc: MultipleChoiceDoc) -> typing.List[str]:
+        return [key2num(doc, key) for key in doc.keys]
+
+
+# TODO: Try to come up with a way to do this it at runtime.
+if os.environ["MC_SETTING"] == "freeform":
+    MULTIPLE_CHOICE_TASK = MC_NoOptionList_OptionLL_Task
+elif os.environ["MC_SETTING"] == "option":
+    MULTIPLE_CHOICE_TASK = MC_WithOptionList_OptionLL_Task
+elif os.environ["MC_SETTING"] == "letter":
+    MULTIPLE_CHOICE_TASK = MC_WithOptionList_LetterLL_Task
+elif os.environ["MC_SETTING"] == "number":
+    MULTIPLE_CHOICE_TASK = MC_WithOptionList_NumLL_Task
+else:
+    print("No such MC_SETTING:", os.environ["MC_SETTING"])
\ No newline at end of file

lm_eval/models/gpt3.py

@@ -66,12 +66,13 @@ def oa_completion(**kwargs):
         except openai.error.OpenAIError:
             import traceback
             traceback.print_exc()
+            traceback.print_exc(file=os.path.join(os.environ["QUESTION_RESULT_PATH"], "traceback.txt"))
             time.sleep(backoff_time)
             backoff_time *= 1.5
 
 
 class GPT3LM(BaseLM):
-    REQ_CHUNK_SIZE = 20
+    REQ_CHUNK_SIZE = 40
 
     def __init__(self, engine, truncate=False, api_key=None, pass_strings=False):
         """
@@ -87,6 +88,7 @@ class GPT3LM(BaseLM):
 
         import openai
         self.engine = engine
+        print(self.max_length)
         self.tokenizer = transformers.GPT2TokenizerFast.from_pretrained('gpt2')
         self.pass_strings = pass_strings
 
@@ -156,7 +158,9 @@ class GPT3LM(BaseLM):
                     inp = self.tok_decode(inp)
                 inps.append(inp)
                 ctxlens.append(ctxlen)
-
+            response = None
+            while True:
+                try:
                     response = oa_completion(
                         engine=self.engine,
                         prompt=inps,
@@ -164,6 +168,12 @@ class GPT3LM(BaseLM):
                         max_tokens=1,
                         logprobs=10,
                     )
+                    break
+                except Exception as e:
+                    print(e)
+                    print("pausing")
+                    time.sleep(1)
+                    continue
 
             for resp, ctxlen, (cache_key, context_enc, continuation_enc) in zip(response.choices, ctxlens, chunk):
                 answer = get_result(resp, ctxlen)
@@ -204,10 +214,14 @@ class GPT3LM(BaseLM):
         for chunk, until in tqdm(list(sameuntil_chunks(reord.get_reordered(), self.REQ_CHUNK_SIZE))):
             inps = []
             for context, _ in chunk:
-                context_enc = self.tok_encode(context)
+                context_enc = self.tok_encode(context, max_length=self.max_length, truncation=False)
                 inp = context_enc[-(self.max_length - self.max_gen_toks):]
                 inps.append(self.tok_decode(inp))
 
+            response = None
+            while True:
+                try:
+
                     response = oa_completion(
                         engine=self.engine,
                         prompt=inps,
@@ -217,6 +231,13 @@ class GPT3LM(BaseLM):
                         stop=until,
                     )
 
+                    break
+                except Exception as e:
+                    print(e)
+                    print("pausing")
+                    time.sleep(1)
+                    continue
+
             for resp, (context, until_) in zip(response.choices, chunk):
                 s = resp['text']
 
@@ -242,7 +263,6 @@ class GPT3LM(BaseLM):
 class GooseAILM(GPT3LM):
     def __init__(self, engine, truncate=False, api_key=None, force_pile_tokenizer=False):
         super().__init__(engine, truncate=truncate, api_key=api_key or os.environ["GOOSEAI_API_SECRET_KEY"], pass_strings=True)
-        self.REQ_CHUNK_SIZE = 1 
         import openai
         openai.api_base = "https://api.goose.ai/v1"
 

lm_eval/tasks/hendrycks_test.py

 import csv
 import random
 from lm_eval.base import MultipleChoiceTask
+from lm_eval.mctask_experimental import MultipleChoiceDoc
 from ..utils import sh
 from pathlib import Path
 from best_download import download_file
@@ -63,26 +64,13 @@ class GeneralHendrycksTest(MultipleChoiceTask):
         return True
 
     def _convert_standard(self, doc):
-        def format_example(doc, choices):
-            """
-                Question: <prompt>
-                Choices:
-                A. <choice1>
-                B. <choice2>
-                C. <choice3>
-                D. <choice4>
-                Answer:
-            """
-            prompt = "Question: " + doc[0] + "\nChoices:\n"
-            prompt += "".join([f"{choices[j]}. {doc[j+1]}\n" for j in range(4)])
-            prompt += "Answer:"
-            return prompt
-        choices = ['A', 'B', 'C', 'D']
-        return {
-            "query": format_example(doc, choices),
-            "choices": doc[1:5],
-            "gold": choices.index(doc[5])
-        }
+        keys = ['A', 'B', 'C', 'D']
+        return MultipleChoiceDoc(
+            question=doc[0],
+            keys=keys,
+            options=doc[1:5],
+            gold=keys.index(doc[5])
+        )
 
     def _load_docs(self, filename):
         reader = csv.reader(open(filename, 'r'), quotechar='"', delimiter=',')
@@ -113,6 +101,3 @@ class GeneralHendrycksTest(MultipleChoiceTask):
             self._fewshot_docs = list(self._load_docs(filename))
 
         return rnd.sample(list(self._fewshot_docs), k)
-
-    def doc_to_text(self, doc):
-        return doc["query"]

mc-answer-prompt-experiment.sh

+# Usage:
+# sh mc-answer-prompt-experiment.sh \
+#   -e <engine> \
+#   -k <number of examples> \
+#   -s <mc-setting = "freeform" | "option" | "letter" | "number"> \
+
+while getopts e:k:s: flag
+do
+    case "${flag}" in
+        e) engine=${OPTARG};;
+        k) k_shot=${OPTARG};;
+        s) setting=${OPTARG};;
+    esac
+done
+
+ENGINE=$engine
+KSHOT=$k_shot
+MC_SETTING=$setting
+
+# Set environment variables.
+#export GOOSEAI_API_SECRET_KEY=sk-
+export MC_SETTING=$setting
+
+# Setup paths.
+RESULT_DIR=$(pwd)/mc-task-results/$ENGINE/$KSHOT-shot
+mkdir -p $RESULT_DIR
+export QUESTION_RESULT_PATH=$RESULT_DIR/$MC_SETTING
+mkdir -p $RESULT_DIR/$MC_SETTING
+
+# Tasks to run.
+HENDRYCKS_TEST=hendrycksTest-abstract_algebra,hendrycksTest-anatomy,hendrycksTest-astronomy,hendrycksTest-business_ethics,hendrycksTest-clinical_knowledge,hendrycksTest-college_biology,hendrycksTest-college_chemistry,hendrycksTest-college_computer_science,hendrycksTest-college_mathematics,hendrycksTest-college_medicine,hendrycksTest-college_physics,hendrycksTest-computer_security,hendrycksTest-conceptual_physics,hendrycksTest-econometrics,hendrycksTest-electrical_engineering,hendrycksTest-elementary_mathematics,hendrycksTest-formal_logic,hendrycksTest-global_facts,hendrycksTest-high_school_biology,hendrycksTest-high_school_chemistry,hendrycksTest-high_school_computer_science,hendrycksTest-high_school_european_history,hendrycksTest-high_school_geography,hendrycksTest-high_school_government_and_politics,hendrycksTest-high_school_macroeconomics,hendrycksTest-high_school_mathematics,hendrycksTest-high_school_microeconomics,hendrycksTest-high_school_physics,hendrycksTest-high_school_psychology,hendrycksTest-high_school_statistics,hendrycksTest-high_school_us_history,hendrycksTest-high_school_world_history,hendrycksTest-human_aging,hendrycksTest-human_sexuality,hendrycksTest-international_law,hendrycksTest-jurisprudence,hendrycksTest-logical_fallacies,hendrycksTest-machine_learning,hendrycksTest-management,hendrycksTest-marketing,hendrycksTest-medical_genetics,hendrycksTest-miscellaneous,hendrycksTest-moral_disputes,hendrycksTest-moral_scenarios,hendrycksTest-nutrition,hendrycksTest-philosophy,hendrycksTest-prehistory,hendrycksTest-professional_accounting,hendrycksTest-professional_law,hendrycksTest-professional_medicine,hendrycksTest-professional_psychology,hendrycksTest-public_relations,hendrycksTest-security_studies,hendrycksTest-sociology,hendrycksTest-us_foreign_policy,hendrycksTest-virology,hendrycksTest-world_religions
+
+# Runner function.
+run_experiment(){
+    local curr_engine=$1
+    local setting=$2
+    local output_path=$RESULT_DIR/$setting
+    
+    # Log stuff.
+    echo "\n"
+    echo "###################################################"
+    echo "PID: $PPID"
+    echo "MC Setting: $setting"
+    echo "Few-shot: $KSHOT"
+    echo "Current Engine: $curr_engine"
+    echo "Current Results Dir:\n$output_path"
+    echo "Start Time: $(date)"
+    echo "###################################################"
+    echo "\n"
+
+    python3 -m scripts.write_out --output_base_path $output_path --tasks hendrycksTest-abstract_algebra --sets test --num_fewshot $KSHOT
+    mv $output_path/hendrycksTest-abstract_algebra $output_path/example_prompt 
+
+    python3 main.py \
+        --model gooseai \
+        --model_args engine=$curr_engine \
+        --tasks $HENDRYCKS_TEST \
+        --output_path $output_path/results.json \
+        --num_fewshot $KSHOT
+
+# Test Call.
+#     python3 main.py \
+#         --device cpu \
+#         --model gpt2 \
+#         --tasks anagrams1 \
+#         --limit 2 \
+#         --output_path $output_path/results.json
+}
+
+# Run experiment.
+touch $RESULT_DIR/$MC_SETTING/out.log
+run_experiment $ENGINE $MC_SETTING > $RESULT_DIR/$MC_SETTING/out.log
+
+# Setup subshells?
+# ()