Update README.md
Browse files
README.md
CHANGED
|
@@ -28,24 +28,106 @@ should probably proofread and complete it, then remove this comment. -->
|
|
| 28 |
|
| 29 |
|
| 30 |
|
| 31 |
-
AlphaMonarch-laser is a DPO fine-tuned of [mlabonne/NeuralMonarch-7B](https://huggingface.co/mlabonne/NeuralMonarch-7B/) using the [argilla/OpenHermes2.5-dpo-binarized-alpha](https://huggingface.co/datasets/argilla/OpenHermes2.5-dpo-binarized-alpha) preference dataset but achieves better performance then [mlabonne/AlphaMonarch-7B](https://huggingface.co/mlabonne/AlphaMonarch-7B/) using
|
| 32 |
|
|
|
|
|
|
|
| 33 |
|
| 34 |
-
|
| 35 |
|
| 36 |
-
#
|
| 37 |
|
| 38 |
-
|
| 39 |
|
| 40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
-
|
| 43 |
|
| 44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
|
| 46 |
-
|
| 47 |
|
| 48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
|
| 50 |
### Training hyperparameters
|
| 51 |
|
|
@@ -63,9 +145,6 @@ The following hyperparameters were used during training:
|
|
| 63 |
|
| 64 |
|
| 65 |
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
### 📝 Axolotl Configuration
|
| 70 |
|
| 71 |
```yaml
|
|
|
|
| 28 |
|
| 29 |
|
| 30 |
|
| 31 |
+
AlphaMonarch-laser is a DPO fine-tuned of [mlabonne/NeuralMonarch-7B](https://huggingface.co/mlabonne/NeuralMonarch-7B/) using the [argilla/OpenHermes2.5-dpo-binarized-alpha](https://huggingface.co/datasets/argilla/OpenHermes2.5-dpo-binarized-alpha) preference dataset but achieves better performance then [mlabonne/AlphaMonarch-7B](https://huggingface.co/mlabonne/AlphaMonarch-7B/) using LaserQLoRA. We have fine-tuned this model only on half of the projections, but have achieved better results as compared to the version released by Maximme Labonne. We have trained this model for 1080 steps.
|
| 32 |
|
| 33 |
+
AlphaMonarch-laser is ranking 1 on YALL - [Yet Another LLM Leaderboard](https://huggingface.co/spaces/mlabonne/Yet_Another_LLM_Leaderboard).
|
| 34 |
+
|
| 35 |
|
| 36 |
+
## 🏆 Evaluation results
|
| 37 |
|
| 38 |
+
# Nous Benchmark
|
| 39 |
|
| 40 |
+
### AGIEVAL
|
| 41 |
|
| 42 |
+
| Task | Version | Metric | Value | StdErr |
|
| 43 |
+
|---------------------------------|---------|--------------|--------|--------|
|
| 44 |
+
| agieval_aqua_rat | 0 | acc | 28.35% | 2.83% |
|
| 45 |
+
| agieval_aqua_rat | 0 | acc_norm | 26.38% | 2.77% |
|
| 46 |
+
| agieval_logiqa_en | 0 | acc | 38.25% | 1.91% |
|
| 47 |
+
| agieval_logiqa_en | 0 | acc_norm | 38.10% | 1.90% |
|
| 48 |
+
| agieval_lsat_ar | 0 | acc | 23.91% | 2.82% |
|
| 49 |
+
| agieval_lsat_ar | 0 | acc_norm | 23.48% | 2.80% |
|
| 50 |
+
| agieval_lsat_lr | 0 | acc | 52.75% | 2.21% |
|
| 51 |
+
| agieval_lsat_lr | 0 | acc_norm | 53.92% | 2.21% |
|
| 52 |
+
| agieval_lsat_rc | 0 | acc | 66.91% | 2.87% |
|
| 53 |
+
| agieval_lsat_rc | 0 | acc_norm | 67.29% | 2.87% |
|
| 54 |
+
| agieval_sat_en | 0 | acc | 78.64% | 2.86% |
|
| 55 |
+
| agieval_sat_en | 0 | acc_norm | 78.64% | 2.86% |
|
| 56 |
+
| agieval_sat_en_without_passage | 0 | acc | 45.15% | 3.48% |
|
| 57 |
+
| agieval_sat_en_without_passage | 0 | acc_norm | 44.17% | 3.47% |
|
| 58 |
+
| agieval_sat_math | 0 | acc | 33.18% | 3.18% |
|
| 59 |
+
| agieval_sat_math | 0 | acc_norm | 31.36% | 3.14% |
|
| 60 |
+
Average: 28.41%
|
| 61 |
|
| 62 |
+
### GPT4ALL
|
| 63 |
|
| 64 |
+
| Task | Version | Metric | Value | StdErr |
|
| 65 |
+
|--------------|---------|----------|-------|--------|
|
| 66 |
+
| arc_challenge| 0 | acc | 66.30%| ± 1.38%|
|
| 67 |
+
| | | acc_norm | 68.26%| ± 1.36%|
|
| 68 |
+
| arc_easy | 0 | acc | 86.57%| ± 0.70%|
|
| 69 |
+
| | | acc_norm | 80.81%| ± 0.81%|
|
| 70 |
+
| boolq | 1 | acc | 87.16%| ± 0.59%|
|
| 71 |
+
| hellaswag | 0 | acc | 69.60%| ± 0.46%|
|
| 72 |
+
| | | acc_norm | 87.45%| ± 0.33%|
|
| 73 |
+
| openbookqa | 0 | acc | 39.20%| ± 2.19%|
|
| 74 |
+
| | | acc_norm | 49.60%| ± 2.24%|
|
| 75 |
+
| piqa | 0 | acc | 83.03%| ± 0.88%|
|
| 76 |
+
| | | acc_norm | 84.87%| ± 0.84%|
|
| 77 |
+
| winogrande | 0 | acc | 81.06%| ± 1.10%|
|
| 78 |
+
Average: 76.98%
|
| 79 |
|
| 80 |
+
### TRUTHFUL-QA
|
| 81 |
|
| 82 |
+
| Task | Version | Metric | Value | StdErr |
|
| 83 |
+
|---------------|---------|--------|-------|--------|
|
| 84 |
+
| truthfulqa_mc | 1 | mc1 | 63.04%| ± 1.69%|
|
| 85 |
+
| truthfulqa_mc | 1 | mc2 | 78.39%| ± 1.37%|
|
| 86 |
+
Average: 70.71%
|
| 87 |
+
|
| 88 |
+
### BIGBENCH
|
| 89 |
+
|
| 90 |
+
| Task | Version | Metric | Value | StdErr |
|
| 91 |
+
|------------------------------------------------|---------|-----------------------|-------|--------------------|
|
| 92 |
+
| bigbench_causal_judgement | 0 | multiple_choice_grade| 60.00%| ± 3.56% |
|
| 93 |
+
| bigbench_date_understanding | 0 | multiple_choice_grade| 62.06%| ± 2.53% |
|
| 94 |
+
| bigbench_disambiguation_qa | 0 | multiple_choice_grade| 54.26%| ± 3.11% |
|
| 95 |
+
| bigbench_geometric_shapes | 0 | multiple_choice_grade| 23.96%| ± 2.26% |
|
| 96 |
+
| | | exact_str_match | 0.00% | ± 0.00% |
|
| 97 |
+
| bigbench_logical_deduction_five_objects | 0 | multiple_choice_grade| 32.80%| ± 2.10% |
|
| 98 |
+
| bigbench_logical_deduction_seven_objects | 0 | multiple_choice_grade| 23.86%| ± 1.61% |
|
| 99 |
+
| bigbench_logical_deduction_three_objects | 0 | multiple_choice_grade| 59.33%| ± 2.84% |
|
| 100 |
+
| bigbench_movie_recommendation | 0 | multiple_choice_grade| 58.00%| ± 2.21% |
|
| 101 |
+
| bigbench_navigate | 0 | multiple_choice_grade| 56.00%| ± 1.57% |
|
| 102 |
+
| bigbench_reasoning_about_colored_objects | 0 | multiple_choice_grade| 69.20%| ± 1.03% |
|
| 103 |
+
| bigbench_ruin_names | 0 | multiple_choice_grade| 55.36%| ± 2.35% |
|
| 104 |
+
| bigbench_salient_translation_error_detection | 0 | multiple_choice_grade| 41.48%| ± 1.56% |
|
| 105 |
+
| bigbench_snarks | 0 | multiple_choice_grade| 73.48%| ± 3.29% |
|
| 106 |
+
| bigbench_sports_understanding | 0 | multiple_choice_grade| 76.06%| ± 1.36% |
|
| 107 |
+
| bigbench_temporal_sequences | 0 | multiple_choice_grade| 55.50%| ± 1.57% |
|
| 108 |
+
| bigbench_tracking_shuffled_objects_five_objects| 0 | multiple_choice_grade| 23.28%| ± 1.20% |
|
| 109 |
+
| bigbench_tracking_shuffled_objects_seven_objects| 0 | multiple_choice_grade| 19.37%| ± 0.94% |
|
| 110 |
+
| bigbench_tracking_shuffled_objects_three_objects| 0 | multiple_choice_grade| 59.33%| ± 2.84% |
|
| 111 |
+
Average: 55.37%
|
| 112 |
+
|
| 113 |
+
# Openllm Benchmark
|
| 114 |
+
|
| 115 |
+
| Task |Version| Metric |Value| |Stderr|
|
| 116 |
+
|-------------|------:|--------|----:|---|-----:|
|
| 117 |
+
|arc_challenge| 0|acc |70.12|± | 1.30|
|
| 118 |
+
| | |acc_norm|73.27|± | 1.29|
|
| 119 |
+
|hellaswag | 0|acc |71.80|± | 0.44|
|
| 120 |
+
| | |acc_norm|89.20|± | 0.30|
|
| 121 |
+
|gsm8k | 0|acc |66.77|± | 1.2 |
|
| 122 |
+
|winogrande | 0|acc |84.6 |± | 1.0 |
|
| 123 |
+
|
| 124 |
+
Average: 73.5%
|
| 125 |
+
|
| 126 |
+
### TruthfulQA
|
| 127 |
+
| Task |Version|Metric|Value| |Stderr|
|
| 128 |
+
|-------------|------:|------|----:|---|-----:|
|
| 129 |
+
|truthfulqa_mc| 1|mc1 |62.79|± | 1.69|
|
| 130 |
+
| | |mc2 |77.90|± | 1.37|
|
| 131 |
|
| 132 |
### Training hyperparameters
|
| 133 |
|
|
|
|
| 145 |
|
| 146 |
|
| 147 |
|
|
|
|
|
|
|
|
|
|
| 148 |
### 📝 Axolotl Configuration
|
| 149 |
|
| 150 |
```yaml
|