Update README.md
Browse files
README.md
CHANGED
|
@@ -42,7 +42,83 @@ The WANDB Project is public and can be examined at this link: https://wandb.ai/t
|
|
| 42 |
|
| 43 |
## Benchmark Information
|
| 44 |
|
| 45 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
|
| 47 |
## Training procedure
|
| 48 |
|
|
|
|
| 42 |
|
| 43 |
## Benchmark Information
|
| 44 |
|
| 45 |
+
## Benchmark Results
|
| 46 |
+
|
| 47 |
+
GPT-4All Benchmark Set
|
| 48 |
+
```
|
| 49 |
+
| Task |Version| Metric |Value | |Stderr|
|
| 50 |
+
|-------------|------:|--------|-----:|---|-----:|
|
| 51 |
+
|arc_challenge| 0|acc |0.5009|± |0.0146|
|
| 52 |
+
| | |acc_norm|0.5247|± |0.0146|
|
| 53 |
+
|arc_easy | 0|acc |0.8127|± |0.0080|
|
| 54 |
+
| | |acc_norm|0.7854|± |0.0084|
|
| 55 |
+
|boolq | 1|acc |0.8153|± |0.0068|
|
| 56 |
+
|hellaswag | 0|acc |0.6126|± |0.0049|
|
| 57 |
+
| | |acc_norm|0.7995|± |0.0040|
|
| 58 |
+
|openbookqa | 0|acc |0.3660|± |0.0216|
|
| 59 |
+
| | |acc_norm|0.4600|± |0.0223|
|
| 60 |
+
|piqa | 0|acc |0.7922|± |0.0095|
|
| 61 |
+
| | |acc_norm|0.8112|± |0.0091|
|
| 62 |
+
|winogrande | 0|acc |0.7293|± |0.0125|
|
| 63 |
+
```
|
| 64 |
+
AGI-Eval
|
| 65 |
+
```
|
| 66 |
+
| Task |Version| Metric |Value | |Stderr|
|
| 67 |
+
|------------------------------|------:|--------|-----:|---|-----:|
|
| 68 |
+
|agieval_aqua_rat | 0|acc |0.2008|± |0.0252|
|
| 69 |
+
| | |acc_norm|0.2126|± |0.0257|
|
| 70 |
+
|agieval_logiqa_en | 0|acc |0.3410|± |0.0186|
|
| 71 |
+
| | |acc_norm|0.3564|± |0.0188|
|
| 72 |
+
|agieval_lsat_ar | 0|acc |0.2261|± |0.0276|
|
| 73 |
+
| | |acc_norm|0.2174|± |0.0273|
|
| 74 |
+
|agieval_lsat_lr | 0|acc |0.3725|± |0.0214|
|
| 75 |
+
| | |acc_norm|0.3373|± |0.0210|
|
| 76 |
+
|agieval_lsat_rc | 0|acc |0.4684|± |0.0305|
|
| 77 |
+
| | |acc_norm|0.4572|± |0.0304|
|
| 78 |
+
|agieval_sat_en | 0|acc |0.6553|± |0.0332|
|
| 79 |
+
| | |acc_norm|0.5971|± |0.0343|
|
| 80 |
+
|agieval_sat_en_without_passage| 0|acc |0.4515|± |0.0348|
|
| 81 |
+
| | |acc_norm|0.4029|± |0.0343|
|
| 82 |
+
|agieval_sat_math | 0|acc |0.3273|± |0.0317|
|
| 83 |
+
| | |acc_norm|0.2636|± |0.0298|
|
| 84 |
+
```
|
| 85 |
+
BigBench Reasoning Test
|
| 86 |
+
```
|
| 87 |
+
| Task |Version| Metric |Value | |Stderr|
|
| 88 |
+
|------------------------------------------------|------:|---------------------|-----:|---|-----:|
|
| 89 |
+
|bigbench_causal_judgement | 0|multiple_choice_grade|0.5368|± |0.0363|
|
| 90 |
+
|bigbench_date_understanding | 0|multiple_choice_grade|0.7127|± |0.0236|
|
| 91 |
+
|bigbench_disambiguation_qa | 0|multiple_choice_grade|0.3023|± |0.0286|
|
| 92 |
+
|bigbench_geometric_shapes | 0|multiple_choice_grade|0.1003|± |0.0159|
|
| 93 |
+
| | |exact_str_match |0.0000|± |0.0000|
|
| 94 |
+
|bigbench_logical_deduction_five_objects | 0|multiple_choice_grade|0.2720|± |0.0199|
|
| 95 |
+
|bigbench_logical_deduction_seven_objects | 0|multiple_choice_grade|0.1986|± |0.0151|
|
| 96 |
+
|bigbench_logical_deduction_three_objects | 0|multiple_choice_grade|0.4500|± |0.0288|
|
| 97 |
+
|bigbench_movie_recommendation | 0|multiple_choice_grade|0.2880|± |0.0203|
|
| 98 |
+
|bigbench_navigate | 0|multiple_choice_grade|0.5000|± |0.0158|
|
| 99 |
+
|bigbench_reasoning_about_colored_objects | 0|multiple_choice_grade|0.5390|± |0.0111|
|
| 100 |
+
|bigbench_ruin_names | 0|multiple_choice_grade|0.3906|± |0.0231|
|
| 101 |
+
|bigbench_salient_translation_error_detection | 0|multiple_choice_grade|0.1844|± |0.0123|
|
| 102 |
+
|bigbench_snarks | 0|multiple_choice_grade|0.5249|± |0.0372|
|
| 103 |
+
|bigbench_sports_understanding | 0|multiple_choice_grade|0.5335|± |0.0159|
|
| 104 |
+
|bigbench_temporal_sequences | 0|multiple_choice_grade|0.2980|± |0.0145|
|
| 105 |
+
|bigbench_tracking_shuffled_objects_five_objects | 0|multiple_choice_grade|0.2048|± |0.0114|
|
| 106 |
+
|bigbench_tracking_shuffled_objects_seven_objects| 0|multiple_choice_grade|0.1297|± |0.0080|
|
| 107 |
+
|bigbench_tracking_shuffled_objects_three_objects| 0|multiple_choice_grade|0.4500|± |0.0288|
|
| 108 |
+
```
|
| 109 |
+
|
| 110 |
+
This is a slight improvement on GPT4ALL Suite and BigBench Suite, with a degredation in AGIEval compared to the original hermes.
|
| 111 |
+
|
| 112 |
+
Average Score Comparison between Nous-Hermes Llama-2 and OpenHermes Llama-2:
|
| 113 |
+
```
|
| 114 |
+
| Bench | Nous-Hermes | OpenHermes | Change |
|
| 115 |
+
|------------------------------|------------:|------------|--------|
|
| 116 |
+
|GPT4All | 70.00| 70.36| +0.36|
|
| 117 |
+
|------------------------------------------------------------------|
|
| 118 |
+
|BigBench | 36.57| 36.75| +0.18|
|
| 119 |
+
|------------------------------------------------------------------|
|
| 120 |
+
|AGI Eval | 37.20| 35.56| -1.64|
|
| 121 |
+
```
|
| 122 |
|
| 123 |
## Training procedure
|
| 124 |
|