File size: 13,824 Bytes
1e59da5 9fd0a82 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 |
---
license: apache-2.0
inference: false
---
# Description
4 bit quantization of [upstage/SOLAR-10.7B-Instruct-v1.0](https://huggingface.co/upstage/SOLAR-10.7B-Instruct-v1.0) using GPTQ. We use the config below for quantization/evaluation and [HuggingFaceH4/ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) as the calibration data. The code is available under [this repository.](https://github.com/IST-DASLab/marlin/tree/2f6d7c10e124b3c5fa29ff8d77d568bd7af3274c/gptq)
```yaml
bits: 4
damp_percent: 0.01
desc_act: true
exllama_config:
version: 2
group_size: 128
quant_method: gptq
static_groups: false
sym: true
true_sequential: true
```
## Evaluations
Below is a comprehensive evaluation using the awesome [mosaicml/llm-foundry](https://github.com/mosaicml/llm-foundry/tree/main/scripts/eval).
| model_name | core_average | world_knowledge | commonsense_reasoning | language_understanding | symbolic_problem_solving | reading_comprehension |
|:----------------------------------|---------------:|------------------:|------------------------:|-------------------------:|---------------------------:|------------------------:|
| upstage/SOLAR-10.7B-Instruct-v1.0 | 0.594131 | 0.602579 | 0.600195 | 0.747605 | 0.406245 | 0.614029 |
| Category | Benchmark | Subtask | Accuracy | Number few shot |
| :----------------------- | :--------------------------- | :---------------------------------- | -------: | :-------------- |
| symbolic_problem_solving | gsm8k | | 0.638362 | 0-shot |
| commonsense_reasoning | copa | | 0.84 | 0-shot |
| commonsense_reasoning | commonsense_qa | | 0.841933 | 0-shot |
| commonsense_reasoning | piqa | | 0.818281 | 0-shot |
| commonsense_reasoning | bigbench_strange_stories | | 0.793103 | 0-shot |
| commonsense_reasoning | bigbench_strategy_qa | | 0.66623 | 0-shot |
| language_understanding | lambada_openai | | 0.735882 | 0-shot |
| language_understanding | hellaswag | | 0.855208 | 0-shot |
| reading_comprehension | coqa | | 0.222723 | 0-shot |
| reading_comprehension | boolq | | 0.893884 | 0-shot |
| world_knowledge | triviaqa_sm_sub | | 0.628333 | 3-shot |
| world_knowledge | jeopardy | Average | 0.500792 | 3-shot |
| world_knowledge | | american_history | 0.581114 | 3-shot |
| world_knowledge | | literature | 0.655102 | 3-shot |
| world_knowledge | | science | 0.371849 | 3-shot |
| world_knowledge | | word_origins | 0.271233 | 3-shot |
| world_knowledge | | world_history | 0.624665 | 3-shot |
| world_knowledge | bigbench_qa_wikidata | | 0.669209 | 3-shot |
| world_knowledge | arc_easy | | 0.815657 | 3-shot |
| world_knowledge | arc_challenge | | 0.650171 | 3-shot |
| commonsense_reasoning | siqa | | 0.881781 | 3-shot |
| language_understanding | winograd | | 0.897436 | 3-shot |
| symbolic_problem_solving | bigbench_operators | | 0.595238 | 3-shot |
| reading_comprehension | squad | | 0.626395 | 3-shot |
| symbolic_problem_solving | svamp | | 0.603333 | 5-shot |
| world_knowledge | mmlu | Average | 0.647028 | 5-shot |
| world_knowledge | | abstract_algebra | 0.29 | 5-shot |
| world_knowledge | | anatomy | 0.577778 | 5-shot |
| world_knowledge | | astronomy | 0.710526 | 5-shot |
| world_knowledge | | business_ethics | 0.73 | 5-shot |
| world_knowledge | | clinical_knowledge | 0.701887 | 5-shot |
| world_knowledge | | college_biology | 0.729167 | 5-shot |
| world_knowledge | | college_chemistry | 0.39 | 5-shot |
| world_knowledge | | college_computer_science | 0.5 | 5-shot |
| world_knowledge | | college_mathematics | 0.31 | 5-shot |
| world_knowledge | | college_medicine | 0.66474 | 5-shot |
| world_knowledge | | college_physics | 0.411765 | 5-shot |
| world_knowledge | | computer_security | 0.72 | 5-shot |
| world_knowledge | | conceptual_physics | 0.582979 | 5-shot |
| world_knowledge | | econometrics | 0.473684 | 5-shot |
| world_knowledge | | electrical_engineering | 0.565517 | 5-shot |
| world_knowledge | | elementary_mathematics | 0.470899 | 5-shot |
| world_knowledge | | formal_logic | 0.460317 | 5-shot |
| world_knowledge | | global_facts | 0.33 | 5-shot |
| world_knowledge | | high_school_biology | 0.770968 | 5-shot |
| world_knowledge | | high_school_chemistry | 0.448276 | 5-shot |
| world_knowledge | | high_school_computer_science | 0.71 | 5-shot |
| world_knowledge | | high_school_european_history | 0.830303 | 5-shot |
| world_knowledge | | high_school_geography | 0.848485 | 5-shot |
| world_knowledge | | high_school_government_and_politics | 0.896373 | 5-shot |
| world_knowledge | | high_school_macroeconomics | 0.646154 | 5-shot |
| world_knowledge | | high_school_mathematics | 0.348148 | 5-shot |
| world_knowledge | | high_school_microeconomics | 0.722689 | 5-shot |
| world_knowledge | | high_school_physics | 0.344371 | 5-shot |
| world_knowledge | | high_school_psychology | 0.833028 | 5-shot |
| world_knowledge | | high_school_statistics | 0.523148 | 5-shot |
| world_knowledge | | high_school_us_history | 0.852941 | 5-shot |
| world_knowledge | | high_school_world_history | 0.827004 | 5-shot |
| world_knowledge | | human_aging | 0.713004 | 5-shot |
| world_knowledge | | human_sexuality | 0.755725 | 5-shot |
| world_knowledge | | international_law | 0.768595 | 5-shot |
| world_knowledge | | jurisprudence | 0.796296 | 5-shot |
| world_knowledge | | logical_fallacies | 0.723926 | 5-shot |
| world_knowledge | | machine_learning | 0.508929 | 5-shot |
| world_knowledge | | management | 0.825243 | 5-shot |
| world_knowledge | | marketing | 0.871795 | 5-shot |
| world_knowledge | | medical_genetics | 0.73 | 5-shot |
| world_knowledge | | miscellaneous | 0.814815 | 5-shot |
| world_knowledge | | moral_disputes | 0.736994 | 5-shot |
| world_knowledge | | moral_scenarios | 0.43352 | 5-shot |
| world_knowledge | | nutrition | 0.728758 | 5-shot |
| world_knowledge | | philosophy | 0.700965 | 5-shot |
| world_knowledge | | prehistory | 0.765432 | 5-shot |
| world_knowledge | | professional_accounting | 0.507092 | 5-shot |
| world_knowledge | | professional_law | 0.487614 | 5-shot |
| world_knowledge | | professional_medicine | 0.727941 | 5-shot |
| world_knowledge | | professional_psychology | 0.661765 | 5-shot |
| world_knowledge | | public_relations | 0.718182 | 5-shot |
| world_knowledge | | security_studies | 0.669388 | 5-shot |
| world_knowledge | | sociology | 0.81592 | 5-shot |
| world_knowledge | | us_foreign_policy | 0.89 | 5-shot |
| world_knowledge | | virology | 0.518072 | 5-shot |
| world_knowledge | | world_religions | 0.789474 | 5-shot |
| symbolic_problem_solving | bigbench_dyck_languages | | 0.458 | 5-shot |
| language_understanding | winogrande | | 0.826361 | 5-shot |
| symbolic_problem_solving | agi_eval_lsat_ar | | 0.269565 | 5-shot |
| symbolic_problem_solving | simple_arithmetic_nospaces | | 0.372 | 5-shot |
| symbolic_problem_solving | simple_arithmetic_withspaces | | 0.367 | 5-shot |
| reading_comprehension | agi_eval_lsat_rc | | 0.794776 | 5-shot |
| reading_comprehension | agi_eval_lsat_lr | | 0.641176 | 5-shot |
| reading_comprehension | agi_eval_sat_en | | 0.849515 | 5-shot |
| world_knowledge | arc_challenge | | 0.670648 | 25-shot |
| commonsense_reasoning | openbook_qa | | 0.56 | 10-shot |
| language_understanding | hellaswag | | 0.866461 | 10-shot |
| | bigbench_cs_algorithms | | 0.652273 | 10-shot |
| symbolic_problem_solving | bigbench_elementary_math_qa | | 0.392453 | 1-shot | |