LelaStarling-7B Overview
LelaStarling-7B is a 7 billion parameter language model developed by luqmanxyz, created through a strategic merge of two distinct models: SanjiWatsuki/Lelantos-DPO-7B and berkeley-nest/Starling-LM-7B-alpha. This merge was executed using the slerp (spherical linear interpolation) method via LazyMergekit, combining the strengths of both base models.
Key Capabilities & Performance
This model demonstrates solid performance across various benchmarks, as evaluated on the Open LLM Leaderboard. Its average score is 71.45, indicating a balanced capability for general language tasks. Specific benchmark results include:
- AI2 Reasoning Challenge (25-Shot): 67.58
- HellaSwag (10-Shot): 86.33
- MMLU (5-Shot): 64.98
- TruthfulQA (0-shot): 57.73
- Winogrande (5-shot): 80.98
- GSM8k (5-shot): 71.11
These scores suggest proficiency in common sense reasoning, factual recall, and mathematical problem-solving, making it suitable for a range of applications requiring robust language understanding and generation.
When to Use This Model
LelaStarling-7B is a strong candidate for use cases requiring a capable 7B parameter model with a balanced performance profile. Its merged architecture aims to provide a versatile foundation for tasks such as:
- General text generation and completion
- Question answering
- Reasoning-based tasks
- Applications benefiting from a model with good common sense understanding