ColQwen3.5 Optimization Trail contains 776+ MTEB evaluation results from the development of three visual document retrieval models. The dataset, published by athrael-soju, captures the full development process including seeds, ablations, and variants for models using ColBERT-style late interaction with Qwen3.5-VL.
Use Cases
- Analyze MTEB evaluation results across seeds and ablation variants to identify hyperparameter optimization patterns.
- Compare benchmark scores for ColQwen3.5-v1, v2, and v3 models to track performance evolution.
- Study the impact of different soup ratios and merge variants on visual document retrieval metrics.
Strengths
- 776+ documented MTEB evaluation results provide a substantial record of model development.
- Captures the full optimization trail including unsuccessful candidates, not just final models.
- Focuses on three specific visual document retrieval models with 4.5 billion parameters.
Limitations
- Dataset structure is undefined with unknown columns and sample data unavailable.
- Limited to the evaluation results of three specific model families, not a general benchmark.
- Recency of last update (2026) may indicate speculative or forward-looking data.
Provenance
- Source
- Hugging Face dataset published by author athrael-soju.
- Collection Method
- MTEB evaluation results collected during the development and hyperparameter optimization of ColQwen3.5 models.
- Time Range
- null
- Freshness
- Last updated 2026-03-15.
- Geography
- null