TAUR-dev/rankalign-v6-gemma-2-2b-d0.15-e2-hc-b2d-dbl-all-tcs-ln-vlo-fsx-lo0.1

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:2.6BQuant:BF16Ctx Length:8kPublished:Apr 6, 2026Architecture:Transformer Warm

The TAUR-dev/rankalign-v6-gemma-2-2b-d0.15-e2-hc-b2d-dbl-all-tcs-ln-vlo-fsx-lo0.1 is a 2.6 billion parameter language model fine-tuned from Google's Gemma-2-2b base model. This model is specifically optimized for hypernym-concat-bananas-to-dogs-double-all tasks, focusing on semantic relationship understanding. It features a context length of 8192 tokens and incorporates typicality correction and length normalization during its training process. Its primary application is in specialized semantic evaluation tasks, particularly those involving hypernym identification.

Loading preview...

Model Overview

This model, rankalign-v6-gemma-2-2b-d0.15-e2-hc-b2d-dbl-all-tcs-ln-vlo-fsx-lo0.1, is a fine-tuned checkpoint derived from the rankalign project, built upon the google/gemma-2-2b base model. With approximately 2.6 billion parameters and an 8192-token context window, it is designed for specific semantic understanding tasks.

Key Training Details

  • Base Model: google/gemma-2-2b
  • Version: v6, trained for 2 epochs with a delta of 0.15.
  • Task Focus: Specifically fine-tuned for hypernym-concat-bananas-to-dogs-double-all tasks, indicating an emphasis on identifying and processing hypernymic relationships.
  • Optimization Features: Incorporates self-typicality correction, length normalization, and validator log-odds during training, suggesting an approach to improve the quality and consistency of generated outputs.
  • Loss Configuration: Utilizes a preference loss weight of 1, with NLL validator and generator weights set to 0, focusing on preference-based learning.

Intended Use Cases

This model is particularly suited for research and development in:

  • Hypernym Identification: Excels in tasks requiring the recognition and generation of hypernyms across various semantic categories.
  • Semantic Relationship Analysis: Useful for exploring and evaluating models' understanding of hierarchical semantic structures.
  • Reproducibility Studies: The provided evaluation scripts allow for direct reproducibility of its performance on a range of hypernym-related tasks, such as hypernym-bananas, hypernym-dogs, and hypernym-elephants.

Finetunes
1 models