AlphaMonarch-dora
AlphaMonarch-dora is a DPO fine-tuned of mlabonne/NeuralMonarch-7B using the 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 and AlphaMonarch-laser. I have trained this model for 1080 steps. All hyperparams were kept consist across all these experiments.
🏆 Evaluation results
OpenLLM Benchmark
Nous Benchmark
Thanks to Muhammad Bin Usman for evaluating AlphaMonarch-DoRA on the NOUS benchmark.
AGIEVAL
| Task | Version | Accuracy | Accuracy StdErr | Normalized Accuracy | Normalized Accuracy StdErr |
|---|---|---|---|---|---|
| agieval_aqua_rat | 0 | 28.35% | 2.83% | 26.38% | 2.77% |
| agieval_logiqa_en | 0 | 38.71% | 1.91% | 38.25% | 1.90% |
| agieval_lsat_ar | 0 | 23.91% | 2.82% | 23.48% | 2.80% |
| agieval_lsat_lr | 0 | 52.55% | 2.21% | 53.73% | 2.21% |
| agieval_lsat_rc | 0 | 66.91% | 2.87% | 66.54% | 2.88% |
| agieval_sat_en | 0 | 78.64% | 2.86% | 78.64% | 2.86% |
| agieval_sat_en_without_passage | 0 | 45.15% | 3.48% | 44.17% | 3.47% |
| agieval_sat_math | 0 | 33.64% | 3.19% | 31.82% | 3.15% |
AVG = 45.976
GPT4ALL
| Task | Version | Accuracy | Accuracy StdErr | Normalized Accuracy | Normalized Accuracy StdErr |
|---|---|---|---|---|---|
| arc_challenge | 0 | 65.87% | 1.39% | 67.92% | 1.36% |
| arc_easy | 0 | 86.49% | 0.70% | 80.64% | 0.81% |
| boolq | 1 | 87.16% | 0.59% | - | - |
| hellaswag | 0 | 69.86% | 0.46% | 87.51% | 0.33% |
| openbookqa | 0 | 39.00% | 2.18% | 49.20% | 2.24% |
| piqa | 0 | 83.03% | 0.88% | 84.82% | 0.84% |
| winogrande | 0 | 80.98% | 1.10% | - | - |
AVG = 73.18
TRUTHFUL-QA
| Task | Version | MC1 Accuracy | MC1 Accuracy StdErr | MC2 Accuracy | MC2 Accuracy StdErr |
|---|---|---|---|---|---|
| truthfulqa_mc | 1 | 62.91% | 1.69% | 78.48% | 1.37% |
AVG = 70.69
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-7
- train_batch_size: 2
- eval_batch_size: Not specified
- seed: Not specified
- gradient_accumulation_steps: 8
- total_train_batch_size: Not specified
- optimizer: PagedAdamW with 32-bit precision
- lr_scheduler_type: Cosine
- lr_scheduler_warmup_steps: 100
- training_steps: 1080
Framework versions
- Transformers 4.39.0.dev0
- Peft 0.9.1.dev0
- Datasets 2.18.0
- torch 2.2.0
- accelerate 0.27.2