Aika-7B
Aika is a language model constructed using the DARE TIES merge method using mitultiwari/mistral-7B-instruct-dpo as a base. Aika is designed to interact with users in a way that feels natural and human-like, to solve problems and answer questions with a high degree of accuracy and truthfulness, and to engage in creative and logical tasks with proficiency.
Models Merged
The following models were included in the merge:
The base model is Mistral-7Bv0.1 fine tuned on Anthropic/hh-rlhf.
Why?
- Base model tuned on Anthropic RLHF dataset: Safe AI as a base model, to balance the uncensored model below.
- Silicon-Maid-7B: Boasts excellent multi-turn conversational skills and logical coherence, ensuring smooth interactions.
- Samantha-V2: Offers empathy and human-like responses, equipped with programmed "self-awareness" for a more personalized experience.
- Stealth-V1.3: Known for enhancing performance in merges when integrated as a component, optimizing Aika's functionality.
- WestLake-7B-V2: Sets a high benchmark for emotional intelligence (EQ) and excels in creative writing, enhancing Aika's ability to understand and respond to your needs.
Combine them all
You get Aika - a considerate, personal digital assistant.
Configuration
Please check mergekit_config.yml for the merge config.
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 59.25 |
| AI2 Reasoning Challenge (25-Shot) | 65.36 |
| HellaSwag (10-Shot) | 81.49 |
| MMLU (5-Shot) | 53.91 |
| TruthfulQA (0-shot) | 51.22 |
| Winogrande (5-shot) | 77.74 |
| GSM8k (5-shot) | 25.78 |
- Downloads last month
- 2,222
Merge of
Dataset used to train sethuiyer/Aika-7B
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard65.360
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard81.490
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard53.910
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard51.220
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard77.740
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard25.780