TAUR-dev/rankalign-v6-gemma-2-9b-it-d0.15-e2-ambigqa-all-tcs-fsx-lo0.1
TAUR-dev/rankalign-v6-gemma-2-9b-it-d0.15-e2-ambigqa-all-tcs-fsx-lo0.1 is a 9 billion parameter language model fine-tuned from Google's Gemma-2-9B-IT base model. Developed as part of the rankalign project, this version (v6) is specifically optimized for the 'ambigqa-all' task. It incorporates a delta of 0.15 and was trained for 2 epochs, making it suitable for tasks requiring nuanced understanding and generation based on ambiguous questions.
Loading preview...
Model Overview
This model, rankalign-v6-gemma-2-9b-it-d0.15-e2-ambigqa-all-tcs-fsx-lo0.1, is a fine-tuned checkpoint derived from the rankalign project. It is built upon the google/gemma-2-9b-it base model, featuring 9 billion parameters.
Key Training Details
- Base Model:
google/gemma-2-9b-it - Version: v6
- Task:
ambigqa-all(indicating a focus on ambiguous question answering) - Epochs: 2
- Delta: 0.15
- Typicality Correction: Self-correction mechanism employed
- Labeled-only Ratio: 0.1
This specific configuration suggests an emphasis on refining the model's ability to handle and respond to ambiguous queries, potentially through preference learning or alignment techniques as implied by the 'rankalign' project origin. The training parameters, including the delta and labeled-only ratio, indicate a targeted fine-tuning approach to enhance performance on complex, ambiguous question-answering scenarios.
Evaluation
The README provides a clear command for evaluating the model's performance on the ambigqa-all task, utilizing scripts/eval_by_claude.py with specific parameters for self-typicality and score saving.