Add GSM8K

backend-cli.py

@@ -28,6 +28,8 @@ import time
 import pprint
 import logging
 
+from lm_eval.filters.extraction import RegexFilter
+
 
 # Configure the root logger
 logging.basicConfig(
@@ -42,6 +44,20 @@ eval_logger = logging.getLogger("lm-eval")
 # Explicitly set the level for 'lm-eval' logger to WARNING
 eval_logger.setLevel(logging.WARNING)
 
+def tuple_input_decorator(func):
+    def wrapper(self, resps, docs):
+        stripped_resps = [[resp_data[0] for resp_data in group] for group in resps]
+
+        filtered_resps = func(self, stripped_resps, docs)
+        
+        combined_resps = []
+        for original_group, new_group in zip(resps, filtered_resps):
+            combined_group = [(new_resp,) + rest_of_data[1:] for new_resp, rest_of_data in zip(new_group, original_group)]
+            combined_resps.append(combined_group)
+
+        return combined_resps
+    return wrapper
+
 
 def my_set_eval_request(api, eval_request, set_to_status, hf_repo, local_dir):
     for i in range(10):
@@ -137,6 +153,12 @@ def process_evaluation(task: Task, eval_request: EvalRequest, limit: Optional[in
     stop_event = threading.Event()
     monitor_thread = threading.Thread(target=monitor_gpus, args=(stop_event, 5, gpu_stats_list))
     monitor_thread.start()
+    
+    original_apply = RegexFilter.apply
+    if task.benchmark == "gsm8k":
+        RegexFilter.apply = tuple_input_decorator(RegexFilter.apply)
+    else:
+        RegexFilter.apply = original_apply
 
     try:
         results = run_evaluation(
@@ -443,7 +465,7 @@ def get_args():
     parser = argparse.ArgumentParser(description="Run the backend")
     parser.add_argument("--debug", action="store_true", help="Run in debug mode")
     # debug parameters
-    parser.add_argument("--task", type=str, default="selfcheckgpt,mmlu", help="Task to debug")
+    parser.add_argument("--task", type=str, default="selfcheckgpt,mmlu, gsm8k", help="Task to debug")
     parser.add_argument("--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1,mistralai/Mixtral-8x7B-v0.1", help="Model to debug")
     parser.add_argument("--precision", type=str, default="float32,float16,8bit,4bit", help="Precision to debug")
     parser.add_argument("--inference-framework", type=str, default="hf-chat", help="Inference framework to debug")
@@ -471,23 +493,23 @@ if __name__ == "__main__":
                     task_name = task.benchmark
                     if task_name not in debug_task_name:
                         continue
-                    try:
-                        eval_request = EvalRequest(
-                            model=debug_model_name,
-                            private=False,
-                            status="",
-                            json_filepath="",
-                            precision=precision,  # Use precision from arguments
-                            inference_framework=args.inference_framework,  # Use inference framework from arguments
-                            gpu_type=args.gpu_type
-                        )
-                        curr_gpu_type = get_gpu_details()
-                        if eval_request.gpu_type != curr_gpu_type:
-                            print(f"GPU type mismatch: {eval_request.gpu_type} vs {curr_gpu_type}")
-                            raise Exception("GPU type mismatch")
-                        results = process_evaluation(task, eval_request, limit=args.limit)
-                    except Exception as e:
-                        print(f"debug running error: {e}")
+                    # try:
+                    eval_request = EvalRequest(
+                        model=debug_model_name,
+                        private=False,
+                        status="",
+                        json_filepath="",
+                        precision=precision,  # Use precision from arguments
+                        inference_framework=args.inference_framework,  # Use inference framework from arguments
+                        gpu_type=args.gpu_type
+                    )
+                    curr_gpu_type = get_gpu_details()
+                    if eval_request.gpu_type != curr_gpu_type:
+                        print(f"GPU type mismatch: {eval_request.gpu_type} vs {curr_gpu_type}")
+                        raise Exception("GPU type mismatch")
+                    results = process_evaluation(task, eval_request, limit=args.limit)
+                    # except Exception as e:
+                    #     print(f"debug running error: {e}")
     else:
         while True:
             res = False

src/backend/envs.py

@@ -57,6 +57,7 @@ class Tasks(Enum):
 
     # task20 = Task("race", "acc", "RACE", 0)
     task21 = Task("mmlu", "acc", "MMLU", 5)
+    task22 = Task("gsm8k", "exact_match", "GSM8K", 5)
 
 
 EVAL_REQUESTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-queue-bk")

src/backend/tasks/gsm8k/README.md

@@ -0,0 +1,59 @@
+# GSM8k
+
+## Paper
+Training Verifiers to Solve Math Word Problems
+https://arxiv.org/abs/2110.14168
+
+State-of-the-art language models can match human performance on many tasks, but
+they still struggle to robustly perform multi-step mathematical reasoning. To
+diagnose the failures of current models and support research, we introduce GSM8K,
+a dataset of 8.5K high quality linguistically diverse grade school math word problems.
+We find that even the largest transformer models fail to achieve high test performance,
+despite the conceptual simplicity of this problem distribution.
+
+NOTE: See the official implementation of the task:
+    https://github.com/openai/grade-school-math/blob/master/grade_school_math/calculator.py
+for how to make use of the dataset's calculator annotations in your language
+model's sample/generation function.
+
+Homepage: https://github.com/openai/grade-school-math
+
+
+## Citation
+```
+@misc{cobbe2021training,
+      title={Training Verifiers to Solve Math Word Problems},
+      author={Karl Cobbe and Vineet Kosaraju and Mohammad Bavarian and Jacob Hilton and Reiichiro Nakano and Christopher Hesse and John Schulman},
+      year={2021},
+      eprint={2110.14168},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG}
+}
+```
+
+### Groups and Tasks
+
+#### Groups
+
+- `math_word_problems`
+- `chain_of_thought`
+- `self_consistency`
+
+#### Tasks
+
+- `gsm8k_yaml`
+- `gsm8k_cot`: GSM8K with Chain-of-Thought
+- `gsm8k_cot_self_consistency`: GSM8K with Chain-of-Thought and Self-Consistency
+
+### Checklist
+
+- [x] Is in Eval-harness v1.0 ?
+- [ ] Has been checked for regression from v1.0?
+- [ ] Has been checked for equivalence with original paper methodology?
+- [ ] "Main" checked variant clearly denoted?
+
+### Variant Wishlist
+
+- [ ] Variant with Calculator (see https://github.com/openai/grade-school-math/blob/master/grade_school_math/calculator.py for example implementation)
+- [ ] Using Verifiers
+- [ ] Majority voting "without CoT"

src/backend/tasks/gsm8k/gsm8k-cot-self-consistency.yaml

@@ -0,0 +1,34 @@
+include: gsm8k-cot.yaml
+group:
+  - chain_of_thought
+  - self_consistency
+task: gsm8k_cot_self_consistency
+generation_kwargs:
+  until:
+    - "Q:"
+    - "\n\n"
+  do_sample: true
+  temperature: 0.2
+repeats: 64
+filter_list:
+  - name: "score-first" # pick only the first response, and report metrics on that
+    filter:
+      - function: "regex"
+        regex_pattern: "The answer is (\\-?[0-9\\.\\,]*[0-9]+)"
+      - function: "take_first"
+  - name: "maj@64"
+    filter:
+      - function: "regex"
+        regex_pattern: "The answer is (\\-?[0-9\\.\\,]*[0-9]+)"
+      - function: "majority_vote"
+      - function: "take_first"
+  - name: "maj@8" # get Maj@8 , via selecting the first 8 responses. Using a better estimator would be optimal.
+    filter:
+      - function: "take_first_k"
+        k: 8
+      - function: "regex"
+        regex_pattern: "The answer is (\\-?[0-9\\.\\,]*[0-9]+)"
+      - function: "majority_vote"
+      - function: "take_first"
+metadata:
+  version: 2.0

src/backend/tasks/gsm8k/gsm8k-cot-zeroshot.yaml

@@ -0,0 +1,44 @@
+group:
+  - math_word_problems
+task: gsm8k_cot_zeroshot
+dataset_path: gsm8k
+dataset_name: main
+output_type: generate_until
+training_split: train
+fewshot_split: train
+test_split: test
+doc_to_text: "Q: {{question}}\nA: Let's think step by step."
+doc_to_target: "{{answer}}" #" {{answer.split('### ')[-1].rstrip()}}"
+metric_list:
+  - metric: exact_match
+    aggregation: mean
+    higher_is_better: true
+    ignore_case: true
+    ignore_punctuation: false
+    regexes_to_ignore:
+      - ","
+      - "\\$"
+      - "(?s).*#### "
+      - "\\.$"
+generation_kwargs:
+  until:
+    - "Q:"
+    - "</s>"
+    - "<|im_end|>"
+  do_sample: false
+repeats: 1
+num_fewshot: 0
+filter_list:
+  - name: "strict-match"
+    filter:
+      - function: "regex"
+        regex_pattern: "The answer is (\\-?[0-9\\.\\,]+)."
+      - function: "take_first"
+  - name: "flexible-extract"
+    filter:
+      - function: "regex"
+        group_select: -1
+        regex_pattern: "(-?[$0-9.,]{2,})|(-?[0-9]+)"
+      - function: "take_first"
+metadata:
+  version: 3.0

src/backend/tasks/gsm8k/gsm8k-cot.yaml

@@ -0,0 +1,51 @@
+group:
+  - chain_of_thought
+task: gsm8k_cot
+dataset_path: gsm8k
+dataset_name: main
+output_type: generate_until
+test_split: test
+doc_to_text: "Q: There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done, there will be 21 trees. How many trees did the grove workers plant today?\nA: There are 15 trees originally. Then there were 21 trees after some more were planted. So there must have been 21 - 15 = 6. The answer is 6.\n\n\
+Q: If there are 3 cars in the parking lot and 2 more cars arrive, how many cars are in the parking lot?\nA: There are originally 3 cars. 2 more cars arrive. 3 + 2 = 5. The answer is 5.\n\n\
+Q: Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?\nA: Originally, Leah had 32 chocolates. Her sister had 42. So in total they had 32 + 42 = 74. After eating 35, they had 74 - 35 = 39. The answer is 39.\n\n\
+Q: Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops. How many lollipops did Jason give to Denny?\nA: Jason started with 20 lollipops. Then he had 12 after giving some to Denny. So he gave Denny 20 - 12 = 8. The answer is 8.\n\n\
+Q: Shawn has five toys. For Christmas, he got two toys each from his mom and dad. How many toys does he have now?\nA: Shawn started with 5 toys. If he got 2 toys each from his mom and dad, then that is 4 more toys. 5 + 4 = 9. The answer is 9.\n\n\
+Q: There were nine computers in the server room. Five more computers were installed each day, from monday to thursday. How many computers are now in the server room?\nA: There were originally 9 computers. For each of 4 days, 5 more computers were added. So 5 * 4 = 20 computers were added. 9 + 20 is 29. The answer is 29.\n\n\
+Q: Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How many golf balls did he have at the end of wednesday?\nA: Michael started with 58 golf balls. After losing 23 on tuesday, he had 58 - 23 = 35. After losing 2 more, he had 35 - 2 = 33 golf balls. The answer is 33.\n\n\
+Q: Olivia has $23. She bought five bagels for $3 each. How much money does she have left?\nA: Olivia had 23 dollars. 5 bagels for 3 dollars each will be 5 x 3 = 15 dollars. So she has 23 - 15 dollars left. 23 - 15 is 8. The answer is 8.\n\n\
+Q: {{question}}\nA:"
+doc_to_target: "{{answer.split('####')[-1].strip()}}"
+metric_list:
+  - metric: exact_match
+    aggregation: mean
+    higher_is_better: true
+    ignore_case: true
+    ignore_punctuation: false
+    regexes_to_ignore:
+      - ","
+      - "\\$"
+      - "(?s).*#### "
+      - "\\.$"
+generation_kwargs:
+  until:
+    - "Q:"
+    - "</s>"
+    - "<|im_end|>"
+  do_sample: false
+repeats: 1
+num_fewshot: 0
+filter_list:
+  - name: "strict-match"
+    filter:
+      - function: "regex"
+        regex_pattern: "The answer is (\\-?[0-9\\.\\,]+)."
+      - function: "take_first"
+  - name: "flexible-extract"
+    filter:
+      - function: "regex"
+        group_select: -1
+        regex_pattern: "(-?[$0-9.,]{2,})|(-?[0-9]+)"
+      - function: "take_first"
+metadata:
+  version: 3.0
+  num_fewshot: 8

src/backend/tasks/gsm8k/gsm8k.yaml

@@ -0,0 +1,45 @@
+group:
+  - math_word_problems
+task: gsm8k
+dataset_path: gsm8k
+dataset_name: main
+output_type: generate_until
+training_split: train
+fewshot_split: train
+test_split: test
+doc_to_text: "Question: {{question}}\nAnswer:"
+doc_to_target: "{{answer}}" #" {{answer.split('### ')[-1].rstrip()}}"
+metric_list:
+  - metric: exact_match
+    aggregation: mean
+    higher_is_better: true
+    ignore_case: true
+    ignore_punctuation: false
+    regexes_to_ignore:
+      - ","
+      - "\\$"
+      - "(?s).*#### "
+      - "\\.$"
+generation_kwargs:
+  until:
+    - "Question:"
+    - "</s>"
+    - "<|im_end|>"
+  do_sample: false
+  temperature: 0.0
+repeats: 1
+num_fewshot: 5
+filter_list:
+  - name: "strict-match"
+    filter:
+      - function: "regex"
+        regex_pattern: "#### (\\-?[0-9\\.\\,]+)"
+      - function: "take_first"
+  - name: "flexible-extract"
+    filter:
+      - function: "regex"
+        group_select: -1
+        regex_pattern: "(-?[$0-9.,]{2,})|(-?[0-9]+)"
+      - function: "take_first"
+metadata:
+  version: 3.0

src/utils.py

@@ -52,7 +52,7 @@ def parse_nvidia_smi():
             print("Failed to query GPU indices.")
             return []
         gpu_indices = result.stdout.strip().split('\n')
-    print(f"gpu_indices: {gpu_indices}")
+    # print(f"gpu_indices: {gpu_indices}")
     gpu_stats = []
 
     gpu_info_pattern = re.compile(r'(\d+)C\s+P\d+\s+(\d+)W / \d+W\s+\|\s+(\d+)MiB / \d+MiB\s+\|\s+(\d+)%')
@@ -80,7 +80,7 @@ def parse_nvidia_smi():
 
             if len(gpu_info) >= 4:
                 gpu_stats.append(gpu_info)
-    print(f"gpu_stats: {gpu_stats}")
+    # print(f"gpu_stats: {gpu_stats}")
     gpu_name = f"{len(gpu_stats)}x{gpu_name}"
     gpu_stats_total = {
                         GPU_TEMP: 0,