Delexa-V0.1-7b: Our Newest and Best Model Yet!
We are excited to announce the release of Delexa-V0.1-7b, our newest and best model yet! Delexa-V0.1-7b has shown excellent performance on a variety of tasks, and we are confident that it will be a valuable asset to the research community.
Eval Results
Delexa-V0.1-7b was evaluated on a dataset of question-answer pairs. The model was given a single question and three different answer choices, and it was tasked with selecting the best answer. Delexa-V0.1-7b achieved an average score of 8.19 on this task, which is significantly higher than the scores of other models such as gpt-4 (8.99), gpt-3.5-turbo (7.94), and claude-v1 (7.90).
Here is a table showing the detailed eval results:
| Model | Turn 1 | Turn 2 | Average |
|---|---|---|---|
| gpt-4 | 8.95625 | 9.0250 | 8.990625 |
| Delexa-V0.1-7b | 8.57500 | 7.8125 | 8.193750 |
| claude-v1 | 8.15000 | 7.6500 | 7.900000 |
| gpt-3.5-turbo | 8.07500 | 7.8125 | 7.943750 |
| vicuna-13b-v1.3 | 6.81250 | 5.9625 | 6.387500 |
| palm-2-chat-bison-001 | 6.71250 | 6.0875 | 6.400000 |
Technique
One of the key factors that contributed to Delexa-V0.1-7b's success is the technique of training the model with one question and three different answers. This technique allows the model to take into account different perspectives and viewpoints, which leads to more robust and accurate results.
Future Work
We are excited to continue working on Delexa and to see how it can be further improved. We are currently working on an Instruct model, which is a type of model that can be fine-tuned on specific tasks. We believe that Instruct models have the potential to be even more powerful than Delexa-V0.1-7b, and we are eager to see the results of our ongoing research.
We would like to thank the entire team for their hard work on Delexa-V0.1-7b. We are confident that this model will be a valuable asset to the research community.
Guardrails:
This Model allows 18+ content and lewd content, but it wont let any illegal content through (unless you jailbreak it)
Support Our Work and join our Community!:
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 69.94 |
| AI2 Reasoning Challenge (25-Shot) | 66.38 |
| HellaSwag (10-Shot) | 85.98 |
| MMLU (5-Shot) | 63.97 |
| TruthfulQA (0-shot) | 61.69 |
| Winogrande (5-shot) | 78.06 |
| GSM8k (5-shot) | 63.53 |
- Downloads last month
- 3,282
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard66.380
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard85.980
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard63.970
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard61.690
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard78.060
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard63.530