llmEval/llm_eval/mmlu.py

209 lines
7.1 KiB
Python

import csv
import os
import re
from typing import Any, Dict, List, Optional
import requests
from llm_eval.base import BaseEvaluator, EvalResult
from llm_eval.registry import EvalFactory
@EvalFactory.register("mmlu")
class MMLUEvaluator(BaseEvaluator):
"""MMLU-style multiple-choice evaluator via HTTP API.
Sends each question as a chat completion request and parses the
answer letter from the response. Two scoring modes:
* ``logprobs`` (default) — requests per-token logprobs and picks the
highest-probability letter among A/B/C/D.
* ``generation`` — asks the model to reply with a single letter and
extracts it from the generated text.
"""
LETTERS = ["A", "B", "C", "D"]
def __init__(
self,
api_base: str,
api_key: str = "not-needed",
model: str = "default",
subject: str = "all",
mode: str = "logprobs",
max_retries: int = 3,
**kwargs,
):
super().__init__(api_base, api_key=api_key)
self.model = model
self.subject = subject
self.mode = mode
self.max_retries = max_retries
# ------------------------------------------------------------------
# Data
# ------------------------------------------------------------------
def load_data(self, data_path: str) -> List[Dict[str, Any]]:
items = []
if self.subject == "all":
for fname in sorted(os.listdir(data_path)):
if fname.endswith(".csv"):
items.extend(self._load_csv(os.path.join(data_path, fname), fname[:-4]))
else:
items = self._load_csv(os.path.join(data_path, f"{self.subject}.csv"), self.subject)
return items
@staticmethod
def _load_csv(path: str, subject: str) -> List[Dict[str, Any]]:
items = []
with open(path, newline="", encoding="utf-8") as f:
reader = csv.reader(f)
for row in reader:
if len(row) < 6:
continue
items.append(dict(
question=row[0],
choices=row[1:5],
answer=ord(row[5].strip().upper()) - ord("A"),
subject=subject,
))
return items
# ------------------------------------------------------------------
# Prompt helpers
# ------------------------------------------------------------------
@staticmethod
def _build_messages(item: Dict[str, Any]) -> List[Dict[str, str]]:
ch = item["choices"]
prompt = (
f"{item['question']}\n\n"
f"A. {ch[0]}\nB. {ch[1]}\nC. {ch[2]}\nD. {ch[3]}"
)
return [
{"role": "system", "content": (
"Answer the multiple-choice question by responding with "
"only the single letter (A, B, C, or D) of the correct answer."
)},
{"role": "user", "content": prompt},
]
# ------------------------------------------------------------------
# API call
# ------------------------------------------------------------------
def _chat(self, messages: List[Dict[str, str]]) -> Dict[str, Any]:
body = dict(model=self.model, messages=messages, temperature=0.0)
if self.mode == "logprobs":
body["logprobs"] = True
body["top_logprobs"] = 5
body["max_tokens"] = 1
else:
body["max_tokens"] = 5
for attempt in range(self.max_retries):
try:
resp = requests.post(
f"{self.api_base}/v1/chat/completions",
json=body,
headers={"Authorization": f"Bearer {self.api_key}"},
timeout=60,
)
resp.raise_for_status()
return resp.json()
except requests.RequestException as e:
if attempt == self.max_retries - 1:
raise
import time
time.sleep(1)
# ------------------------------------------------------------------
# Scoring
# ------------------------------------------------------------------
@staticmethod
def _extract_letter(text: str) -> Optional[int]:
m = re.search(r"\b([A-D])\b", text.strip().upper())
return ord(m.group(1)) - ord("A") if m else None
def _score_logprobs(self, data: List[Dict[str, Any]]) -> List[int]:
preds = []
for item in data:
resp = self._chat(self._build_messages(item))
choices = resp.get("choices", [])
if not choices:
preds.append(-1)
continue
lp_data = choices[0].get("logprobs")
if not lp_data or not lp_data.get("top_logprobs"):
text = choices[0].get("message", {}).get("content", "")
preds.append(self._extract_letter(text) or -1)
continue
top = lp_data["top_logprobs"][0]
best = -1
best_lp = float("-inf")
for letter in self.LETTERS:
found = False
for token, lp in top.items():
if token.strip().upper() == letter:
if lp > best_lp:
best_lp = lp
best = ord(letter) - ord("A")
found = True
break
if not found:
continue
if best == -1:
text = choices[0].get("message", {}).get("content", "")
preds.append(self._extract_letter(text) or -1)
else:
preds.append(best)
return preds
def _score_generation(self, data: List[Dict[str, Any]]) -> List[int]:
preds = []
for item in data:
resp = self._chat(self._build_messages(item))
choices = resp.get("choices", [])
text = choices[0]["message"]["content"] if choices else ""
preds.append(self._extract_letter(text) or -1)
return preds
# ------------------------------------------------------------------
# Evaluate
# ------------------------------------------------------------------
def evaluate(self, data_path: str) -> EvalResult:
data = self.load_data(data_path)
scorer = self._score_logprobs if self.mode == "logprobs" else self._score_generation
preds = scorer(data)
correct = 0
results = []
for item, pred in zip(data, preds):
ok = pred == item["answer"]
correct += int(ok)
results.append(dict(
subject=item["subject"],
question=item["question"],
choices=item["choices"],
answer=item["answer"],
prediction=pred,
correct=ok,
))
total = len(data)
return EvalResult(
task_name=f"mmlu_{self.subject}",
num_samples=total,
accuracy=correct / total if total else 0.0,
results=results,
metadata={"subject": self.subject, "mode": self.mode, "model": self.model},
)