Merge branch 'main' into longcxt

lm_eval/models/huggingface.py

@@ -818,6 +818,12 @@ class HFLM(TemplateLM):
             **add_special_tokens,
         )
         if left_truncate_len:
+            original_lengths = encoding["input_ids"].size(1)
+            if original_lengths > left_truncate_len:
+                eval_logger.warn(
+                    f"Left truncation applied. Original sequence length was {original_lengths}, "
+                    f"truncating to last {left_truncate_len} tokens. Some content will be lost.",
+                )
             encoding["input_ids"] = encoding["input_ids"][:, -left_truncate_len:]
             encoding["attention_mask"] = encoding["attention_mask"][
                 :, -left_truncate_len:
@@ -1096,6 +1102,13 @@ class HFLM(TemplateLM):
 
                 # when too long to fit in context, truncate from the left
                 if self.backend == "causal":
+                    total_length = len(context_enc) + len(continuation_enc)
+                    if total_length > self.max_length + 1:
+                        eval_logger.warn(
+                            f"Combined length of context ({len(context_enc)}) and continuation ({len(continuation_enc)}) "
+                            f"exceeds model's maximum length ({self.max_length}). "
+                            f"Truncating {total_length - self.max_length + 1} tokens from the left."
+                        )
                     inp = torch.tensor(
                         (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                         dtype=torch.long,
@@ -1303,6 +1316,9 @@ class HFLM(TemplateLM):
             if self.backend == "causal":
                 # max len for inputs = max length, minus room to generate the max new tokens
                 max_ctx_len = self.max_length - max_gen_toks
+                assert (
+                    max_ctx_len > 0
+                ), f"Invalid configuration: requested max tokens to generate ({max_gen_toks}) must be less than model's maximum sequence length ({self.max_length})."
             elif self.backend == "seq2seq":
                 # max len for inputs = encoder's whole max_length
                 max_ctx_len = self.max_length

lm_eval/tasks/README.md

@@ -45,6 +45,7 @@
 | [fld](fld/README.md) | Tasks involving free-form and directed dialogue understanding. | English |
 | [french_bench](french_bench/README.md) | Set of tasks designed to assess language model performance in French. | French|
 | [galician_bench](galician_bench/README.md) | Collection of tasks in Galician encompassing various evaluation areas. | Galician |
+| [global_mmlu](global_mmlu/README.md) | Collection of culturally sensitive and culturally agnostic MMLU tasks in 15 languages with human translations or post-edits. | Multiple (15 languages) |
 | [glue](glue/README.md) | General Language Understanding Evaluation benchmark to test broad language abilities. | English |
 | [gpqa](gpqa/README.md) | Tasks designed for general public question answering and knowledge verification. | English |
 | [gsm8k](gsm8k/README.md) | A benchmark of grade school math problems aimed at evaluating reasoning capabilities. | English |

lm_eval/tasks/global_mmlu/README.md

+# Global-MMLU
+
+### Paper
+
+Title: `Global MMLU: Understanding and Addressing Cultural and Linguistic Biases in Multilingual Evaluation`
+
+Abstract: [https://arxiv.org/abs/2412.03304](https://arxiv.org/abs/2412.03304)
+
+Global-MMLU-Lite is a balanced collection of culturally sensitive and culturally agnostic MMLU tasks. It is designed for efficient evaluation of multilingual models in 15 languages (including English). Only languages with human translations and post-edits in the original [Global-MMLU](https://huggingface.co/datasets/CohereForAI/Global-MMLU) 🌍 dataset have been included in the lite version.
+
+Homepage: [https://huggingface.co/datasets/CohereForAI/Global-MMLU-Lite](https://huggingface.co/datasets/CohereForAI/Global-MMLU-Lite)
+
+### Citation
+
+```bibtex
+@misc{singh2024globalmmluunderstandingaddressing,
+      title={Global MMLU: Understanding and Addressing Cultural and Linguistic Biases in Multilingual Evaluation},
+      author={Shivalika Singh and Angelika Romanou and Clémentine Fourrier and David I. Adelani and Jian Gang Ngui and Daniel Vila-Suero and Peerat Limkonchotiwat and Kelly Marchisio and Wei Qi Leong and Yosephine Susanto and Raymond Ng and Shayne Longpre and Wei-Yin Ko and Madeline Smith and Antoine Bosselut and Alice Oh and Andre F. T. Martins and Leshem Choshen and Daphne Ippolito and Enzo Ferrante and Marzieh Fadaee and Beyza Ermis and Sara Hooker},
+      year={2024},
+      eprint={2412.03304},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2412.03304},
+}
+```

lm_eval/tasks/global_mmlu/_default_yaml

+tag:
+  - global_mmlu
+dataset_path: CohereForAI/Global-MMLU-Lite
+test_split: test
+fewshot_split: dev
+fewshot_config:
+  sampler: default
+output_type: multiple_choice
+doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:"
+doc_to_choice: ["A", "B", "C", "D"]
+doc_to_target: answer
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 0.0

lm_eval/tasks/global_mmlu/_generate_configs.py

+import yaml
+
+
+languages = [
+    "en",
+    "ar",
+    "fr",
+    "es",
+    "hi",
+    "de",
+    "id",
+    "it",
+    "ja",
+    "ko",
+    "pt",
+    "zh",
+    "yo",
+    "bn",
+    "sw",
+]
+
+
+def main() -> None:
+    for language in languages:
+        file_name = f"global_mmlu_{language}.yaml"
+        try:
+            with open(f"{file_name}", "w") as f:
+                f.write("# Generated by _generate_configs.py\n")
+                yaml.dump(
+                    {
+                        "include": "_default_yaml",
+                        "task": f"global_mmlu_{language}",
+                        "dataset_name": language,
+                    },
+                    f,
+                )
+        except FileExistsError:
+            pass
+
+
+if __name__ == "__main__":
+    main()

lm_eval/tasks/global_mmlu/global_mmlu_ar.yaml

+# Generated by _generate_configs.py
+dataset_name: ar
+include: _default_yaml
+task: global_mmlu_ar

lm_eval/tasks/global_mmlu/global_mmlu_bn.yaml

+# Generated by _generate_configs.py
+dataset_name: bn
+include: _default_yaml
+task: global_mmlu_bn

lm_eval/tasks/global_mmlu/global_mmlu_de.yaml

+# Generated by _generate_configs.py
+dataset_name: de
+include: _default_yaml
+task: global_mmlu_de

lm_eval/tasks/global_mmlu/global_mmlu_en.yaml

+# Generated by _generate_configs.py
+dataset_name: en
+include: _default_yaml
+task: global_mmlu_en

lm_eval/tasks/global_mmlu/global_mmlu_es.yaml

+# Generated by _generate_configs.py
+dataset_name: es
+include: _default_yaml
+task: global_mmlu_es

lm_eval/tasks/global_mmlu/global_mmlu_fr.yaml

+# Generated by _generate_configs.py
+dataset_name: fr
+include: _default_yaml
+task: global_mmlu_fr

lm_eval/tasks/global_mmlu/global_mmlu_hi.yaml

+# Generated by _generate_configs.py
+dataset_name: hi
+include: _default_yaml
+task: global_mmlu_hi

lm_eval/tasks/global_mmlu/global_mmlu_id.yaml

+# Generated by _generate_configs.py
+dataset_name: id
+include: _default_yaml
+task: global_mmlu_id

lm_eval/tasks/global_mmlu/global_mmlu_it.yaml

+# Generated by _generate_configs.py
+dataset_name: it
+include: _default_yaml
+task: global_mmlu_it

lm_eval/tasks/global_mmlu/global_mmlu_ja.yaml

+# Generated by _generate_configs.py
+dataset_name: ja
+include: _default_yaml
+task: global_mmlu_ja

lm_eval/tasks/global_mmlu/global_mmlu_ko.yaml

+# Generated by _generate_configs.py
+dataset_name: ko
+include: _default_yaml
+task: global_mmlu_ko

lm_eval/tasks/global_mmlu/global_mmlu_pt.yaml

+# Generated by _generate_configs.py
+dataset_name: pt
+include: _default_yaml
+task: global_mmlu_pt

lm_eval/tasks/global_mmlu/global_mmlu_sw.yaml

+# Generated by _generate_configs.py
+dataset_name: sw
+include: _default_yaml
+task: global_mmlu_sw

lm_eval/tasks/global_mmlu/global_mmlu_yo.yaml

+# Generated by _generate_configs.py
+dataset_name: yo
+include: _default_yaml
+task: global_mmlu_yo

lm_eval/tasks/global_mmlu/global_mmlu_zh.yaml

+# Generated by _generate_configs.py
+dataset_name: zh
+include: _default_yaml
+task: global_mmlu_zh