Merge pull request #249 from eshwarprasadS/add-longbench

Add LongBench V2 benchmark

Author: RobotSail

pyproject.toml

@@ -41,6 +41,7 @@ issues = "https://github.com/instructlab/eval/issues"
 "mt_bench" = "instructlab.eval.mt_bench:MTBenchEvaluator"
 "mt_bench_branch" = "instructlab.eval.mt_bench:MTBenchBranchEvaluator"
 "leaderboard_v2" = "instructlab.eval.leaderboard:LeaderboardV2Evaluator"
+"longbench" = "instructlab.eval.longbench:LongBenchEvaluator"
 "ruler" = "instructlab.eval.ruler:RulerEvaluator"
 
 [tool.setuptools_scm]
@@ -53,7 +54,11 @@ package-dir = {"" = "src"}
 
 [tool.setuptools.dynamic]
 dependencies = {file = ["requirements.txt"]}
-optional-dependencies = {leaderboard = {file = ["requirements-leaderboard.txt"]}}
+
+[tool.setuptools.dynamic.optional-dependencies]
+leaderboard = {file = ["requirements-leaderboard.txt"]}
+longbench = {file = ["requirements-longbench.txt"]}
+cuda = {file = ["requirements-cuda.txt"]}
 
 [tool.setuptools.packages.find]
 where = ["src"]

requirements-cuda.txt

@@ -0,0 +1,2 @@
+vllm
+flash-attn

requirements-longbench.txt

@@ -0,0 +1 @@
+lm-eval[longbench]>=0.4.8

src/instructlab/eval/longbench.py

@@ -0,0 +1,306 @@
+# Standard
+from collections import defaultdict
+import json
+import os
+import typing as t
+
+# Third Party
+from lm_eval.evaluator import simple_evaluate
+from torch import cuda
+
+# Local
+from .evaluator import Evaluator
+
+
+class LongBenchResult(t.TypedDict, total=False):
+    """Dict containing averages for each task type and language"""
+
+    overall_score: float
+    en_multidoc: float
+    zh_multidoc: float
+    en_singledoc: float
+    zh_singledoc: float
+    en_summ: float
+    zh_summ: float
+    en_fewshot: float
+    zh_fewshot: float
+    en_synthetic: float
+    zh_synthetic: float
+    code_avg: float
+
+
+# Define task categories
+TASK_CATEGORIES = {
+    "en_multidoc": ["longbench_hotpotqa", "longbench_2wikimqa", "longbench_musique"],
+    "zh_multidoc": ["longbench_dureader"],
+    "en_singledoc": [
+        "longbench_multifieldqa_en",
+        "longbench_narrativeqa",
+        "longbench_qasper",
+    ],
+    "zh_singledoc": ["longbench_multifieldqa_zh"],
+    "en_summ": ["longbench_gov_report", "longbench_qmsum", "longbench_multi_news"],
+    "zh_summ": ["longbench_vcsum"],
+    "en_fewshot": ["longbench_triviaqa", "longbench_samsum", "longbench_trec"],
+    "zh_fewshot": ["longbench_lsht"],
+    "en_synthetic": ["longbench_passage_retrieval_en", "longbench_passage_count"],
+    "zh_synthetic": ["longbench_passage_retrieval_zh"],
+    "code_avg": ["longbench_lcc", "longbench_repobench-p"],
+}
+
+# Flatten the categories to get all tasks
+ALL_LONGBENCH_TASKS = []
+for task in TASK_CATEGORIES.values():
+    ALL_LONGBENCH_TASKS.extend(task)
+
+# Task to metric mapping
+TASK_METRICS = {
+    "longbench_hotpotqa": "qa_f1_score",
+    "longbench_2wikimqa": "qa_f1_score",
+    "longbench_musique": "qa_f1_score",
+    "longbench_dureader": "rouge_zh_score",
+    "longbench_multifieldqa_en": "qa_f1_score",
+    "longbench_narrativeqa": "qa_f1_score",
+    "longbench_qasper": "qa_f1_score",
+    "longbench_multifieldqa_zh": "qa_f1_zh_score",
+    "longbench_gov_report": "rouge_score",
+    "longbench_qmsum": "rouge_score",
+    "longbench_multi_news": "rouge_score",
+    "longbench_vcsum": "rouge_zh_score",
+    "longbench_triviaqa": "qa_f1_score",
+    "longbench_samsum": "rouge_score",
+    "longbench_trec": "classification_score",
+    "longbench_lsht": "classification_score",
+    "longbench_passage_retrieval_en": "retrieval_score",
+    "longbench_passage_count": "count_score",
+    "longbench_passage_retrieval_zh": "retrieval_zh_score",
+    "longbench_lcc": "code_sim_score",
+    "longbench_repobench-p": "code_sim_score",
+}
+
+# Default configuration parameters
+# pylint: disable=use-dict-literal
+DEFAULT_EVAL_CONFIG = dict(
+    batch_size="auto",
+    apply_chat_template=True,
+    fewshot_as_multiturn=True,
+    confirm_run_unsafe_code=True,
+    system_instruction=None,
+    cache_requests=False,
+)
+
+# vLLM-specific configuration - using longer context window than leaderboard
+# pylint: disable=use-dict-literal
+DEFAULT_VLLM_CONFIG = dict(
+    dtype="float16",
+    gpu_memory_utilization=0.8,
+    disable_custom_all_reduce=True,
+    enforce_eager=False,
+    max_model_len=131072,  # 128K context for LongBench
+)
+
+# OpenAI API configuration parameters
+# pylint: disable=use-dict-literal
+DEFAULT_OPENAI_CONFIG = dict(
+    max_tokens=768,
+    temperature=0.0,
+    seed=1337,
+)
+
+
+class LongBenchEvaluator(Evaluator):
+    """
+    Evaluator for LongBench benchmark.
+
+    Attributes:
+        model_path: Path to the model or model name for API
+        tasks: List of subtasks to evaluate (default is all tasks)
+        num_gpus: Number of GPUs to use for local evaluation
+        output_file: Path to save results to
+        eval_config: Configuration for evaluation parameters
+        vllm_config: Configuration for vLLM-specific parameters
+        openai_config: Configuration for OpenAI API parameters
+        api_endpoint: Optional OpenAI-compatible API endpoint
+    """
+
+    name = "longbench"
+
+    def __init__(
+        self,
+        model_path: str,
+        model_name: str,
+        tasks: t.Optional[t.List[str]] = None,
+        num_gpus: t.Optional[int] = None,
+        output_file: t.Optional[str] = None,
+        eval_config: t.Optional[t.Dict[str, t.Any]] = None,
+        vllm_config: t.Optional[t.Dict[str, t.Any]] = None,
+        openai_config: t.Optional[t.Dict[str, t.Any]] = None,
+        api_endpoint: t.Optional[str] = None,
+    ):
+        self.model_path = model_path
+        self.model_name = model_name
+        self.tasks = tasks or ALL_LONGBENCH_TASKS
+
+        # If using API, no need to check CUDA
+        self.api_endpoint = api_endpoint
+        if not api_endpoint and not cuda.is_available():
+            raise ValueError(
+                "Running without CUDA is currently unsupported unless using an API endpoint"
+            )
+
+        self.num_gpus = num_gpus or cuda.device_count()
+        self.output_file = output_file
+        self.eval_config = eval_config if eval_config else {}
+        self.vllm_config = vllm_config if vllm_config else {}
+        self.openai_config = openai_config if openai_config else {}
+        self._results: t.Optional[LongBenchResult] = None
+        self._lm_eval_results: t.Optional[t.Dict[str, t.Any]] = None
+
+    def _get_task_averages(self, results: dict) -> LongBenchResult:
+        """Calculate averages for each task type and language from raw results"""
+        eval_results = defaultdict(float)
+        results_data = results["results"]
+
+        # Track which categories have data
+        active_categories = {}
+
+        # Process each category
+        for category, category_tasks in TASK_CATEGORIES.items():
+            # Filter tasks that were actually run
+            active_tasks = [task for task in category_tasks if task in results_data]
+
+            if active_tasks:
+                # Get scores for active tasks
+                scores = []
+                # pylint: disable=redefined-outer-name
+                for task in active_tasks:
+                    metric_key = f"{TASK_METRICS[task]},none"
+                    if task in results_data and metric_key in results_data[task]:
+                        scores.append(results_data[task][metric_key])
+
+                if scores:
+                    # Calculate average for this category
+                    eval_results[category] = sum(scores) / len(scores)
+                    active_categories[category] = len(scores)
+
+        # Calculate overall score from active categories
+        category_scores = [v for k, v in eval_results.items() if k != "overall_score"]
+        if category_scores:
+            eval_results["overall_score"] = sum(category_scores) / len(category_scores)
+        else:
+            eval_results["overall_score"] = 0.0
+
+        return t.cast(LongBenchResult, dict(eval_results))
+
+    def run(
+        self,
+        model_path: t.Optional[str] = None,
+        model_name: t.Optional[str] = None,
+        tasks: t.Optional[t.List[str]] = None,
+        num_gpus: t.Optional[int] = None,
+        output_file: t.Optional[str] = None,
+        eval_config: t.Optional[t.Dict[str, t.Any]] = None,
+        vllm_config: t.Optional[t.Dict[str, t.Any]] = None,
+        openai_config: t.Optional[t.Dict[str, t.Any]] = None,
+        api_endpoint: t.Optional[str] = None,
+    ) -> LongBenchResult:
+        """Run the LongBench evaluation"""
+        model_path = model_path or self.model_path
+        model_name = model_name or self.model_name
+        tasks = tasks or self.tasks
+        num_gpus = num_gpus or self.num_gpus
+        output_file = output_file or self.output_file
+        api_endpoint = api_endpoint or self.api_endpoint
+
+        # Merge configurations
+        final_eval_config = {}
+        final_eval_config.update(DEFAULT_EVAL_CONFIG)
+        final_eval_config.update(self.eval_config)
+        if eval_config:
+            final_eval_config.update(eval_config)
+
+        final_vllm_config = {}
+        final_vllm_config.update(DEFAULT_VLLM_CONFIG)
+        final_vllm_config.update(self.vllm_config)
+        if vllm_config:
+            final_vllm_config.update(vllm_config)
+
+        final_openai_config = {}
+        final_openai_config.update(DEFAULT_OPENAI_CONFIG)
+        final_openai_config.update(self.openai_config)
+        if openai_config:
+            final_openai_config.update(openai_config)
+
+        # Extract system_instruction if provided
+        system_instruction = final_eval_config.pop("system_instruction", None)
+
+        # Run evaluation with the appropriate backend
+        if api_endpoint:
+            base_url = api_endpoint
+            api_key = final_openai_config.pop("api_key", None)
+
+            # Build model args
+            model_args = {
+                "model": model_name,
+                "tokenizer": model_path,
+                "base_url": base_url,
+            }
+            # Optionally add max_length
+            if "max_length" in final_openai_config:
+                model_args["max_length"] = str(final_openai_config["max_length"])
+
+            if api_key:
+                model_args["api_key"] = str(api_key)
+
+            # Add any other openai_config keys if needed
+            # model_args.update(final_openai_config)
+
+            # Run evaluation
+            results = simple_evaluate(
+                tasks=tasks,
+                model="local-completions",
+                model_args=model_args,
+                system_instruction=system_instruction,
+                **final_eval_config,
+            )
+        else:
+            # Prepare vLLM model args
+            model_args = {
+                "pretrained": model_path,
+                "data_parallel_size": str(num_gpus),
+            }
+            # Add vllm config properly - convert all values to strings
+            string_vllm_config = {k: str(v) for k, v in final_vllm_config.items()}
+            model_args.update(string_vllm_config)
+
+            # Run evaluation
+            results = simple_evaluate(
+                tasks=tasks,
+                model="vllm",
+                model_args=model_args,
+                system_instruction=system_instruction,
+                **final_eval_config,
+            )
+
+        self._lm_eval_results = results
+        self._results = self._get_task_averages(results)
+
+        if output_file:
+            self.save_to_file(output_file)
+
+        return self._results
+
+    @property
+    def results(self) -> t.Optional[LongBenchResult]:
+        """Returns the results of the most recent evaluation"""
+        return self._results
+
+    def save_to_file(self, output_file: t.Optional[str] = None) -> None:
+        """Save results to a JSON file"""
+        output_file = output_file or self.output_file
+        if not output_file:
+            raise ValueError("Output file path cannot be empty")
+
+        os.makedirs(os.path.dirname(output_file), exist_ok=True)
+        with open(output_file, "w", encoding="utf-8") as f:
+            json.dump(self._results, f, indent=2)

tests/test_project.py

@@ -5,6 +5,7 @@
 # First Party
 from instructlab.eval.evaluator import Evaluator
 from instructlab.eval.leaderboard import LeaderboardV2Evaluator
+from instructlab.eval.longbench import LongBenchEvaluator
 from instructlab.eval.mmlu import MMLUBranchEvaluator, MMLUEvaluator
 from instructlab.eval.mt_bench import MTBenchBranchEvaluator, MTBenchEvaluator
 from instructlab.eval.ruler import RulerEvaluator
@@ -17,6 +18,7 @@ def test_evaluator_eps():
         "mt_bench": MTBenchEvaluator,
         "mt_bench_branch": MTBenchBranchEvaluator,
         "leaderboard_v2": LeaderboardV2Evaluator,
+        "longbench": LongBenchEvaluator,
         "ruler": RulerEvaluator,
     }
     eps = entry_points(group="instructlab.eval.evaluator")