|
| 1 | +# SPDX-License-Identifier: Apache-2.0 |
| 2 | + |
| 3 | +# Standard |
| 4 | +from unittest import mock |
| 5 | + |
| 6 | +# First Party |
| 7 | +from instructlab.eval.mt_bench_common import Judge, check_data |
| 8 | + |
| 9 | +CHECK_DATA_EXAMPLE_QUESTIONS = [ |
| 10 | + { |
| 11 | + "question_id": 81, |
| 12 | + "category": "writing", |
| 13 | + "turns": [ |
| 14 | + "Fake question", |
| 15 | + "Fake question", |
| 16 | + ], |
| 17 | + }, |
| 18 | + { |
| 19 | + "question_id": 101, |
| 20 | + "category": "reasoning", |
| 21 | + "turns": [ |
| 22 | + "Fake question", |
| 23 | + "Fake question", |
| 24 | + ], |
| 25 | + }, |
| 26 | +] |
| 27 | +CHECK_DATA_EXAMPLE_MODEL_ANSWERS = { |
| 28 | + "granite-7b-lab": { |
| 29 | + 81: { |
| 30 | + "question_id": 81, |
| 31 | + "answer_id": "c4j9vPyHM8w3JHPGohrJQG", |
| 32 | + "model_id": "granite-7b-lab", |
| 33 | + "choices": [ |
| 34 | + { |
| 35 | + "index": 0, |
| 36 | + "turns": [ |
| 37 | + "Fake answer", |
| 38 | + "Fake answer", |
| 39 | + ], |
| 40 | + } |
| 41 | + ], |
| 42 | + "tstamp": 1730816201.883507, |
| 43 | + }, |
| 44 | + 101: { |
| 45 | + "question_id": 101, |
| 46 | + "answer_id": "kaQw7Fj2SDeE2VfvU25FJ4", |
| 47 | + "model_id": "granite-7b-lab", |
| 48 | + "choices": [ |
| 49 | + { |
| 50 | + "index": 0, |
| 51 | + "turns": [ |
| 52 | + "Fake answer", |
| 53 | + "Fake answer", |
| 54 | + ], |
| 55 | + } |
| 56 | + ], |
| 57 | + "tstamp": 1730816166.3719094, |
| 58 | + }, |
| 59 | + } |
| 60 | +} |
| 61 | +CHECK_DATA_EXAMPLE_REFERENCE_ANSWERS = { |
| 62 | + "merlinite-7b-lab": { |
| 63 | + 101: { |
| 64 | + "question_id": 101, |
| 65 | + "answer_id": "TFomieEmmAgdeCkvmuvwbc", |
| 66 | + "model_id": "gpt-4", |
| 67 | + "choices": [ |
| 68 | + { |
| 69 | + "index": 0, |
| 70 | + "turns": [ |
| 71 | + "Fake answer", |
| 72 | + "Fake answer", |
| 73 | + ], |
| 74 | + } |
| 75 | + ], |
| 76 | + "tstamp": 1686286924.844282, |
| 77 | + }, |
| 78 | + 102: { |
| 79 | + "question_id": 102, |
| 80 | + "answer_id": "hLH8WozvaB88bb5vV224H4", |
| 81 | + "model_id": "gpt-4", |
| 82 | + "choices": [ |
| 83 | + { |
| 84 | + "index": 0, |
| 85 | + "turns": [ |
| 86 | + "Fake answer", |
| 87 | + "Fake answer", |
| 88 | + ], |
| 89 | + } |
| 90 | + ], |
| 91 | + "tstamp": 1686286937.7164738, |
| 92 | + }, |
| 93 | + } |
| 94 | +} |
| 95 | + |
| 96 | +CHECK_DATA_EXAMPLE_MODELS = ["granite-7b-lab"] |
| 97 | +CHECK_DATA_EXAMPLE_JUDGES = { |
| 98 | + "default": Judge( |
| 99 | + model_name="merlinite-7b-lab", |
| 100 | + prompt_template={ |
| 101 | + "name": "single-v1", |
| 102 | + "type": "single", |
| 103 | + "system_prompt": "Fake prompt", |
| 104 | + "prompt_template": "Fake prompt", |
| 105 | + "description": "Prompt for general questions", |
| 106 | + "category": "general", |
| 107 | + "output_format": "[[rating]]", |
| 108 | + }, |
| 109 | + ref_based=False, |
| 110 | + multi_turn=False, |
| 111 | + ), |
| 112 | + "math": Judge( |
| 113 | + model_name="merlinite-7b-lab", |
| 114 | + prompt_template={ |
| 115 | + "name": "single-math-v1", |
| 116 | + "type": "single", |
| 117 | + "system_prompt": "Fake prompt", |
| 118 | + "prompt_template": "Fake prompt", |
| 119 | + "description": "Prompt for general questions", |
| 120 | + "category": "math", |
| 121 | + "output_format": "[[rating]]", |
| 122 | + }, |
| 123 | + ref_based=True, |
| 124 | + multi_turn=False, |
| 125 | + ), |
| 126 | + "default-mt": Judge( |
| 127 | + model_name="merlinite-7b-lab", |
| 128 | + prompt_template={ |
| 129 | + "name": "single-v1-multi-turn", |
| 130 | + "type": "single", |
| 131 | + "system_prompt": "Fake prompt", |
| 132 | + "prompt_template": "Fake prompt", |
| 133 | + "description": "Prompt for general questions", |
| 134 | + "category": "general", |
| 135 | + "output_format": "[[rating]]", |
| 136 | + }, |
| 137 | + ref_based=False, |
| 138 | + multi_turn=True, |
| 139 | + ), |
| 140 | + "math-mt": Judge( |
| 141 | + model_name="merlinite-7b-lab", |
| 142 | + prompt_template={ |
| 143 | + "name": "single-math-v1-multi-turn", |
| 144 | + "type": "single", |
| 145 | + "system_prompt": "Fake prompt", |
| 146 | + "prompt_template": "Fake prompt", |
| 147 | + "description": "Prompt for general questions", |
| 148 | + "category": "math", |
| 149 | + "output_format": "[[rating]]", |
| 150 | + }, |
| 151 | + ref_based=True, |
| 152 | + multi_turn=True, |
| 153 | + ), |
| 154 | +} |
| 155 | + |
| 156 | + |
| 157 | +def test_check_data(): |
| 158 | + check_data( |
| 159 | + CHECK_DATA_EXAMPLE_QUESTIONS, |
| 160 | + CHECK_DATA_EXAMPLE_MODEL_ANSWERS, |
| 161 | + CHECK_DATA_EXAMPLE_REFERENCE_ANSWERS, |
| 162 | + CHECK_DATA_EXAMPLE_MODELS, |
| 163 | + CHECK_DATA_EXAMPLE_JUDGES, |
| 164 | + ) |
| 165 | + |
| 166 | + try: |
| 167 | + check_data( |
| 168 | + CHECK_DATA_EXAMPLE_QUESTIONS, |
| 169 | + {"granite-7b-lab": {}}, |
| 170 | + CHECK_DATA_EXAMPLE_REFERENCE_ANSWERS, |
| 171 | + CHECK_DATA_EXAMPLE_MODELS, |
| 172 | + CHECK_DATA_EXAMPLE_JUDGES, |
| 173 | + ) |
| 174 | + except Exception as e: |
| 175 | + assert "Missing model granite-7b-lab's answer to Question" in str(e) |
| 176 | + else: |
| 177 | + assert False, "Didn't fail with missing model answer" |
| 178 | + |
| 179 | + try: |
| 180 | + check_data( |
| 181 | + CHECK_DATA_EXAMPLE_QUESTIONS, |
| 182 | + CHECK_DATA_EXAMPLE_MODEL_ANSWERS, |
| 183 | + {"merlinite-7b-lab": {}}, |
| 184 | + CHECK_DATA_EXAMPLE_MODELS, |
| 185 | + CHECK_DATA_EXAMPLE_JUDGES, |
| 186 | + ) |
| 187 | + except Exception as e: |
| 188 | + assert "Missing reference answer to Question" in str(e) |
| 189 | + else: |
| 190 | + assert False, "Didn't fail with missing reference answer" |
0 commit comments