QueryloopAI/AlphaMonarch-dora
022
1---2license: cc-by-nc-4.03base_model: mlabonne/NeuralMonarch-7B4tags:5- generated_from_trainer6- mistral7- instruct8- finetune9- chatml10- gpt411- synthetic data12- distillation13model-index:14- name: AlphaMonarch-dora15 results: []16datasets:17- argilla/OpenHermes2.5-dpo-binarized-alpha18language:19- en20library_name: transformers21pipeline_tag: text-generation22---23# AlphaMonarch-dora24 2526 27 28 29<!-- Provide a quick summary of what the model is/does. -->30AlphaMonarch-dora 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 using DoRA. This model is slightly less performant on the Nous and Openllm leaderboards in comparison to base [AlphaMonarch](https://huggingface.co/mlabonne/AlphaMonarch-7B) and [AlphaMonarch-laser](https://huggingface.co/abideen/AlphaMonarch-laser). I have trained this model for 1080 steps. All hyperparams were kept consist across all these experiments.31 32 33## 🏆 Evaluation results34 35# OpenLLM Benchmark36 37 3839 40# Nous Benchmark41 42### AGIEVAL43 44| Task | Version | Accuracy | Accuracy StdErr | Normalized Accuracy | Normalized Accuracy StdErr |45|--------------------------------|---------|----------|-----------------|---------------------|-----------------------------|46| agieval_aqua_rat | 0 | 28.35% | 2.83% | 26.38% | 2.77% |47| agieval_logiqa_en | 0 | 38.71% | 1.91% | 38.25% | 1.90% |48| agieval_lsat_ar | 0 | 23.91% | 2.82% | 23.48% | 2.80% |49| agieval_lsat_lr | 0 | 52.55% | 2.21% | 53.73% | 2.21% |50| agieval_lsat_rc | 0 | 66.91% | 2.87% | 66.54% | 2.88% |51| agieval_sat_en | 0 | 78.64% | 2.86% | 78.64% | 2.86% |52| agieval_sat_en_without_passage | 0 | 45.15% | 3.48% | 44.17% | 3.47% |53| agieval_sat_math | 0 | 33.64% | 3.19% | 31.82% | 3.15% |54 55AVG = 45.97656 57### GPT4ALL58 59| Task | Version | Accuracy | Accuracy StdErr | Normalized Accuracy | Normalized Accuracy StdErr |60|--------------|---------|----------|-----------------|---------------------|-----------------------------|61| arc_challenge| 0 | 65.87% | 1.39% | 67.92% | 1.36% |62| arc_easy | 0 | 86.49% | 0.70% | 80.64% | 0.81% |63| boolq | 1 | 87.16% | 0.59% | - | - |64| hellaswag | 0 | 69.86% | 0.46% | 87.51% | 0.33% |65| openbookqa | 0 | 39.00% | 2.18% | 49.20% | 2.24% |66| piqa | 0 | 83.03% | 0.88% | 84.82% | 0.84% |67| winogrande | 0 | 80.98% | 1.10% | - | - |68 69AVG = 73.1870 71### TRUTHFUL-QA72 73| Task | Version | MC1 Accuracy | MC1 Accuracy StdErr | MC2 Accuracy | MC2 Accuracy StdErr |74|---------------|---------|--------------|---------------------|--------------|---------------------|75| truthfulqa_mc | 1 | 62.91% | 1.69% | 78.48% | 1.37% |76 77AVG = 70.6978 79### Training hyperparameters80The following hyperparameters were used during training:81- learning_rate: 5e-782- train_batch_size: 283- eval_batch_size: Not specified84- seed: Not specified85- gradient_accumulation_steps: 886- total_train_batch_size: Not specified87- optimizer: PagedAdamW with 32-bit precision88- lr_scheduler_type: Cosine89- lr_scheduler_warmup_steps: 10090- training_steps: 108091### Framework versions92- Transformers 4.39.0.dev093- Peft 0.9.1.dev094- Datasets 2.18.095- torch 2.2.096- accelerate 0.27.2