Fix LLaMA tokenization issue (#531)

* Fix tokenization issue in BaseLM.loglikelihood

* Add a regression script

* Use entire non-continuation length as context

---------
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

lm_eval/base.py

@@ -170,16 +170,27 @@ class BaseLM(LM):
     # subclass must implement properties vocab_size, eot_token_id, max_gen_toks, batch_size, device, max_length.
     # TODO: enforce this somehow
 
+    def _encode_pair(self, context, continuation):
+        whole_enc = self.tok_encode(context + continuation)
+        whole_enc_len = len(whole_enc)
+        context_enc = self.tok_encode(context)
+        context_enc_len = len(context_enc)
+        if context_enc_len < whole_enc_len:
+            continuation_enc = whole_enc[context_enc_len:]
+        else:
+            continuation_enc = self.tok_encode(continuation)
+            continuation_enc_len = len(continuation_enc)
+            context_enc = whole_enc[:-continuation_enc_len]
+        return context_enc, continuation_enc
+
     def loglikelihood(self, requests):
         new_reqs = []
         for context, continuation in requests:
             if context == "":
                 # end of text as context
-                context_enc = [self.eot_token_id]
+                context_enc, continuation_enc = [self.eot_token_id], self.tok_encode(continuation)
             else:
-                context_enc = self.tok_encode(context)
-
-            continuation_enc = self.tok_encode(continuation)
+                context_enc, continuation_enc = self._encode_pair(context, continuation)
 
             new_reqs.append(((context, continuation), context_enc, continuation_enc))
 

lm_eval/utils.py

@@ -5,6 +5,7 @@ import collections
 import functools
 import inspect
 import sys
+import fnmatch
 from typing import List, Union
 
 import torch
@@ -84,6 +85,42 @@ def group(arr, fn):
     return list(res.values())
 
 
+def _is_json_task(task_name):
+    return task_name == "json" or task_name.startswith("json=")
+
+
+class MultiChoice:
+    def __init__(self, choices):
+        self.choices = choices
+
+    # Simple wildcard support (linux filename patterns)
+    def __contains__(self, values):
+        for value in values.split(","):
+            if len(fnmatch.filter(self.choices, value)) == 0 and not _is_json_task(
+                value
+            ):
+                return False
+
+        return True
+
+    def __iter__(self):
+        for choice in self.choices:
+            yield choice
+
+
+# Returns a list containing all values of the source_list that
+# match at least one of the patterns
+def pattern_match(patterns, source_list):
+    task_names = set()
+    for pattern in patterns:
+        if _is_json_task(pattern):
+            task_names.add(pattern)
+
+        for matching in fnmatch.filter(source_list, pattern):
+            task_names.add(matching)
+    return sorted(list(task_names))
+
+
 def general_detokenize(string):
     string = string.replace(" n't", "n't")
     string = string.replace(" )", ")")

main.py

 import argparse
 import json
 import logging
-import fnmatch
 import os
 
-from lm_eval import tasks, evaluator
+from lm_eval import tasks, evaluator, utils
 
 logging.getLogger("openai").setLevel(logging.WARNING)
 
 
-def _is_json_task(task_name):
-    return task_name == "json" or task_name.startswith("json=")
-
-
-class MultiChoice:
-    def __init__(self, choices):
-        self.choices = choices
-
-    # Simple wildcard support (linux filename patterns)
-    def __contains__(self, values):
-        for value in values.split(","):
-            if len(fnmatch.filter(self.choices, value)) == 0 and not _is_json_task(
-                value
-            ):
-                return False
-
-        return True
-
-    def __iter__(self):
-        for choice in self.choices:
-            yield choice
-
-
 def parse_args():
     parser = argparse.ArgumentParser()
     parser.add_argument("--model", required=True)
     parser.add_argument("--model_args", default="")
-    parser.add_argument("--tasks", default=None, choices=MultiChoice(tasks.ALL_TASKS))
+    parser.add_argument("--tasks", default=None, choices=utils.MultiChoice(tasks.ALL_TASKS))
     parser.add_argument("--provide_description", action="store_true")
     parser.add_argument("--num_fewshot", type=int, default=0)
     parser.add_argument("--batch_size", type=str, default=None)
@@ -56,19 +32,6 @@ def parse_args():
     return parser.parse_args()
 
 
-# Returns a list containing all values of the source_list that
-# match at least one of the patterns
-def pattern_match(patterns, source_list):
-    task_names = set()
-    for pattern in patterns:
-        if _is_json_task(pattern):
-            task_names.add(pattern)
-
-        for matching in fnmatch.filter(source_list, pattern):
-            task_names.add(matching)
-    return sorted(list(task_names))
-
-
 def main():
     args = parse_args()
 
@@ -82,7 +45,7 @@ def main():
     if args.tasks is None:
         task_names = tasks.ALL_TASKS
     else:
-        task_names = pattern_match(args.tasks.split(","), tasks.ALL_TASKS)
+        task_names = utils.pattern_match(args.tasks.split(","), tasks.ALL_TASKS)
 
     print(f"Selected Tasks: {task_names}")
 

scripts/regression.py

+import argparse
+import json
+import os
+import subprocess
+import time
+from pathlib import Path
+
+from lm_eval import tasks, utils
+
+
+seq2seq_models = ["google/flan-t5-small"]
+causal_models = ["gpt2", "facebook/opt-125m", "EleutherAI/gpt-neo-125m", "EleutherAI/pythia-160m"]
+model_names = seq2seq_models + causal_models
+
+
+completion_tasks = ["boolq", "lambada_openai", "winogrande"]
+choice_tasks = ["hellaswag", "openbookqa", "piqa"]
+perplexity_tasks = ["wikitext"]
+generation_tasks = []
+task_names = completion_tasks + choice_tasks + perplexity_tasks + generation_tasks
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--branches", default=[])
+    parser.add_argument("--models", default=model_names)
+    parser.add_argument("--tasks", default=task_names)
+    parser.add_argument("--acc_norm", type=bool, default=False)
+    parser.add_argument("--perplexity", default=None)
+    # TODO: implement num_fewshot and limit per task, e.g. task1:5,task2:1:100,task3::1000
+    parser.add_argument("--num_fewshot", type=int, default=0)
+    parser.add_argument("--limit", type=float, default=None)
+    # TODO: implement hf-auto to pick between causal and seq2seq models so we don't need this
+    parser.add_argument("--model", default="hf-causal-experimental")
+    # Use whatever is faster here
+    parser.add_argument("--model_args", default="use_accelerate=True,load_in_8bit=True")
+    parser.add_argument("--batch_size", default="auto")
+    return parser.parse_args()
+
+
+def eval_models(args, branch=None):
+    if branch is not None:
+        if os.system(f"git checkout {branch}") != 0:
+            return {}, 0
+
+    branch = branch or initial_branch
+
+    start_time = time.time()
+
+    results = {}
+
+    for model in args.models:
+        model_type = "hf-causal-experimental" if model in causal_models \
+            else "hf-seq2seq" if model in seq2seq_models else args.model
+        model_args = f"pretrained={model},{args.model_args}"
+        # TODO: split_and_pad_windows in AutoSeq2SeqLM doesn"t exist, #527
+        tasks = args.tasks if model in causal_models or model_type == "hf-causal-experimental" \
+            else list(filter(lambda task: task not in perplexity_tasks, args.tasks))
+        # TODO: OOM with auto for seq2seq models, also can OOM with llama
+        batch_size = args.batch_size if model in causal_models or model_type == "hf-causal-experimental" \
+            else 64 if args.batch_size == "auto" else args.batch_size
+        output_path = f"data/regression/{int(start_time)}-{branch}-{Path(model).name}.json"
+
+        command = f"python3 main.py --model {model_type} --model_args {model_args} --tasks {','.join(tasks)} " \
+                  f"--num_fewshot {args.num_fewshot}{'' if args.limit is None else f' --limit {args.limit}'} " \
+                  f"--batch_size {batch_size} --no_cache --output_path {output_path}"
+
+        print(f"{'=' * 80}\nEvaluating {model} on {', '.join(tasks)} at {branch} with:\n\n{command}\n{'=' * 80}")
+
+        ret = os.system(command)
+
+        results[model] = json.load(open(output_path)) if ret == 0 else {"results": {}}
+
+    end_time = time.time()
+
+    return results, end_time - start_time
+
+
+def extract_value(args, results, model, task, err=False):
+    if model not in results:
+        return 0
+    results = results[model]["results"]
+    if task not in results:
+        return 0
+    results = results[task]
+    if args.acc_norm and "acc_norm" in results:
+        return results["acc_norm"] if not err else results["acc_norm_stderr"]
+    if "acc" in results:
+        return results["acc"] if not err else results["acc_stderr"]
+    if (args.perplexity or "word_perplexity") in results:
+        return results[args.perplexity or "word_perplexity"] if not err else 0
+    return 0
+
+
+def format_value(args, results, model, task):
+    val = 100 * extract_value(args, results, model, task)
+    err = 100 * extract_value(args, results, model, task, err=True)
+    return f"{val:.2f}{f' ± {err:.2f}' if err != 0 else ''}"
+
+
+def format_diff(args, results1, results2, model, task):
+    val1 = 100 * extract_value(args, results1, model, task)
+    val2 = 100 * extract_value(args, results2, model, task)
+    diff = val2 - val1
+    return f"**+{diff:.2f}**" if diff > 0 else f"{diff:.2f}"
+
+
+def main():
+    args = parse_args()
+
+    args.branches = args.branches.split(",") if type(args.branches) == str else args.branches
+    args.models = args.models.split(",") if type(args.models) == str else args.models
+    args.tasks = tasks.ALL_TASKS if args.tasks == "all_tasks" \
+        else utils.pattern_match(args.tasks.split(",") if type(args.tasks) == str else args.tasks, tasks.ALL_TASKS)
+
+    global initial_branch
+    initial_branch = subprocess.check_output("git branch --show-current", shell=True).decode("ascii").strip()
+
+    # TODO: implement proper timing for each task
+    # TODO: reduce IO by sharing tasks between models?
+
+    results, runtime = eval_models(args)
+    print(results, runtime)
+
+    runs = []
+    for branch in args.branches:
+        runs.append((branch, *eval_models(args, branch)))
+
+    os.system(f"git checkout {initial_branch}")
+
+    print("")
+    print(f"|task|{'|'.join(map(lambda model: Path(model).name, args.models))}|")
+    print(f"|--|{'--|' * len(args.models)}")
+    for task in args.tasks:
+        print(f"|{task} ({initial_branch})|{'|'.join(map(lambda model: format_value(args, results, model, task), args.models))}|")
+        for branch, branch_results, branch_runtime in runs:
+            print(f"|{task} ({branch})|{'|'.join(map(lambda model: format_value(args, branch_results, model, task), args.models))}|")
+            print(f"|{task} (diff)|{'|'.join(map(lambda model: format_diff(args, results, branch_results, model, task), args.models))}|")
+
+    print("")
+    print("|branch|runtime|%|")
+    print("|--|--|--|")
+    print(f"|{initial_branch}|{runtime:.1f}s|100%|")
+    for branch, _, branch_runtime in runs:
+        print(f"|{branch}|{branch_runtime:.1f}s|{100 * branch_runtime / runtime:.2f}%|")
+
+
+if __name__ == "__main__":
+    main()