add llama3 tasks

lm_eval/tasks/llama3/README.md

@@ -29,6 +29,9 @@ BibTeX-formatted citation goes here
 
 #### Tasks
 
+### Tasks
+
+* `mgsm_chat`: 0-shot mgsm benchmark. Use with chat-template.
 * `mmlu_llama`: `generation variant of MMLU`
 * `arc_chalenge_chat`: `generation variant of ARC-Challenge using MMLU format`
 

lm_eval/tasks/llama3/base/arc_challenge.yaml

@@ -0,0 +1,24 @@
+tag:
+  - llama
+task: llama_arc_challenge
+dataset_path: allenai/ai2_arc
+dataset_name: ARC-Challenge
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+test_split: test
+fewshot_split: train
+doc_to_text: "Question: {{question.strip()}}\nA. {{choices.text[0]}}\nB. {{choices.text[1]}}\nC. {{choices.text[2]}}{% if choices.text|length > 3 %}\nD. {{choices.text[3]}}{% endif %}\nAnswer:"
+fewshot_delimiter: "\n\n"
+doc_to_target: "{{ 'ABCD'[answerKey|int - 1] if answerKey|string in '1234' else answerKey }}"
+doc_to_choice: "{{ choices.label|map('replace', '1', 'A')|map('replace', '2', 'B')|map('replace', '3', 'C')|map('replace', '4', 'D')|list if choices.label[0] in '1234' else choices.label }}"
+num_fewshot: 25
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/llama3/base/utils.py

@@ -0,0 +1,15 @@
+import datasets
+
+
+def process_arc_c_docs(dataset: datasets.Dataset) -> datasets.Dataset:
+    COLUMNS = dataset.column_names
+
+    def map_(doc):
+        doc["doc_to_text"] = doc["input_final_prompts"][0].strip()[:-2].strip()
+        doc["doc_to_choice"] = [
+            x.replace("Answer:", "").strip() for x in doc["output_choice_completions"]
+        ]
+        doc["doc_to_target"] = doc["input_correct_responses"][0].strip()[-1]
+        return doc
+
+    return dataset.map(map_, remove_columns=COLUMNS)

lm_eval/tasks/llama3/instruct/math/README.md

@@ -0,0 +1,66 @@
+# MATH
+ℹ️ This is the 0-shot variant! reproducing https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__math__details?row=0
+## Paper
+Measuring Mathematical Problem Solving With the MATH Dataset
+https://arxiv.org/abs/2103.03874
+
+Many intellectual endeavors require mathematical problem solving, but this skill remains beyond the capabilities of computers. To measure this ability in machine learning models, we introduce MATH, a new dataset of 12,500 challenging competition mathematics problems. Each problem in MATH has a full step-by-step solution which can be used to teach models to generate answer derivations and explanations.
+
+NOTE: The few-shot and the generated answer extraction is based on the [Minerva](https://arxiv.org/abs/2206.14858) and exact match equivalence is calculated using the `sympy` library. This requires additional dependencies, which can be installed via the `lm-eval[math]` extra.
+
+Homepage: https://github.com/hendrycks/math
+
+
+## Citation
+```
+@article{hendrycksmath2021,
+  title={Measuring Mathematical Problem Solving With the MATH Dataset},
+  author={Dan Hendrycks and Collin Burns and Saurav Kadavath and Akul Arora and Steven Basart and Eric Tang and Dawn Song and Jacob Steinhardt},
+  journal={NeurIPS},
+  year={2021}
+}
+
+@misc{2206.14858,
+Author = {Aitor Lewkowycz and Anders Andreassen and David Dohan and Ethan Dyer and Henryk Michalewski and Vinay Ramasesh and Ambrose Slone and Cem Anil and Imanol Schlag and Theo Gutman-Solo and Yuhuai Wu and Behnam Neyshabur and Guy Gur-Ari and Vedant Misra},
+Title = {Solving Quantitative Reasoning Problems with Language Models},
+Year = {2022},
+Eprint = {arXiv:2206.14858},
+}
+```
+
+### Groups and Tasks
+
+[//]: # (#### Groups)
+
+[//]: # ()
+[//]: # (- `llama_math`)
+
+#### Tasks
+
+- `llama_math_algebra`
+- `llama_math_counting_and_prob`
+- `llama_math_geometry`
+- `llama_math_intermediate_algebra`
+- `llama_math_num_theory`
+- `llama_math_prealgebra`
+- `llama_math_precalc`
+
+### Checklist
+
+The checklist is the following:
+
+For adding novel benchmarks/datasets to the library:
+* [x] Is the task an existing benchmark in the literature?
+  * [x] Have you referenced the original paper that introduced the task?
+  * [x] If yes, does the original paper provide a reference implementation? If so, have you checked against the reference implementation and documented how to run such a test?
+    * The implementation in the original paper is one where the model is first fine-tuned on the data. They do have a few-shot evaluation for GPT-3, however the few-shot context used here is sourced from [Lewkowycz et al](https://arxiv.org/abs/2206.14858). The achieved accuracy on Llama-2 models is comparable to that provided in the paper, though not identical.
+
+
+If other tasks on this dataset are already supported:
+* [x] Is the "Main" variant of this task clearly denoted?
+* [x] Have you provided a short sentence in a README on what each new variant adds / evaluates?
+* [x] Have you noted which, if any, published evaluation setups are matched by this variant?
+
+### Variant Wishlist
+
+- [ ] zero-shot variant

lm_eval/tasks/llama3/instruct/math/llama_math_algebra.yaml

@@ -0,0 +1,25 @@
+task: llama_math_algebra
+dataset_path: EleutherAI/hendrycks_math
+process_docs: !function utils.process_docs
+dataset_name: algebra
+output_type: generate_until
+training_split: train
+test_split: test
+doc_to_text: "Solve the following math problem efficiently and clearly:\n\n- For simple problems (2 steps or fewer):\nProvide a concise solution with minimal explanation.\n\n- For complex problems (3 steps or more):\nUse this step-by-step format:\n\n## Step 1: [Concise description]\n[Brief explanation and calculations]\n\n## Step 2: [Concise description]\n[Brief explanation and calculations]\n\n...\n\nRegardless of the approach, always conclude with:\n\nTherefore, the final answer is: $\\\\boxed{answer}$. I hope it is correct.\n\nWhere [answer] is just the final number or expression that solves the problem.\n\nProblem: {{ problem }}"
+process_results: !function utils.process_results
+doc_to_target: "{{answer if few_shot is undefined else solution}}"
+generation_kwargs:
+  until:
+    - "Problem:"
+  max_gen_toks: 5120
+  do_sample: false
+  temperature: 0
+metric_list:
+  - metric: exact_match
+    aggregation: mean
+    higher_is_better: true
+num_fewshot: 0
+metadata:
+  version: 1.0
+dataset_kwargs:
+  trust_remote_code: true

lm_eval/tasks/llama3/instruct/math/llama_math_counting_and_prob.yaml

@@ -0,0 +1,3 @@
+include: llama_math_algebra.yaml
+dataset_name: counting_and_probability
+task: llama_math_counting_and_prob

lm_eval/tasks/llama3/instruct/math/llama_math_geometry.yaml

@@ -0,0 +1,3 @@
+include: llama_math_algebra.yaml
+dataset_name: geometry
+task: llama_math_geometry

lm_eval/tasks/llama3/instruct/math/llama_math_intermediate_algebra.yaml

@@ -0,0 +1,3 @@
+include: llama_math_algebra.yaml
+dataset_name: intermediate_algebra
+task: llama_math_intermediate_algebra

lm_eval/tasks/llama3/instruct/math/llama_math_num_theory.yaml

@@ -0,0 +1,3 @@
+include: llama_math_algebra.yaml
+dataset_name: number_theory
+task: llama_math_num_theory

lm_eval/tasks/llama3/instruct/math/llama_math_prealgebra.yaml

@@ -0,0 +1,3 @@
+include: llama_math_algebra.yaml
+dataset_name: prealgebra
+task: llama_math_prealgebra

lm_eval/tasks/llama3/instruct/math/llama_math_precalc.yaml

@@ -0,0 +1,3 @@
+include: llama_math_algebra.yaml
+dataset_name: precalculus
+task: llama_math_precalc

lm_eval/tasks/llama3/instruct/math/math.yaml

@@ -0,0 +1,14 @@
+group: llama_math
+task:
+  - llama_math_algebra
+  - llama_math_counting_and_prob
+  - llama_math_geometry
+  - llama_math_intermediate_algebra
+  - llama_math_num_theory
+  - llama_math_prealgebra
+  - llama_math_precalc
+aggregate_metric_list:
+  - metric: exact_match
+    weight_by_size: True
+metadata:
+  version: 1