Krystalan/DRT-14B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Dec 23, 2024License:cc-by-nc-sa-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Krystalan/DRT-14B is a 14.8 billion parameter causal language model, fine-tuned from Qwen2.5-14B-Instruct, developed by Krystalan for Deep Reasoning Translation (DRT). This model is specifically designed for neural machine translation (MT) tasks that benefit from long chain-of-thought reasoning, particularly for complex English sentences containing similes or metaphors. It leverages a multi-agent framework for synthesizing MT samples and demonstrates strong performance in translation quality, outperforming its backbone and other models on metrics like GRF, CometKiwi, and BLEU.

Loading preview...

DRT-14B: Deep Reasoning Translation Model

DRT-14B is a 14.8 billion parameter model developed by Krystalan, fine-tuned from the Qwen2.5-14B-Instruct backbone. It is part of the DRT (Deep Reasoning Translation) series, which explores integrating long chain-of-thought reasoning into neural machine translation (MT).

Key Capabilities

  • Long Chain-of-Thought Translation: Specifically designed to handle complex English sentences, such as those with similes or metaphors, by generating an explicit reasoning process (<thought>) before producing the final translation (<output>).
  • Multi-Agent Synthesized Data: Trained on 22,264 synthesized MT samples generated by a multi-agent framework comprising a translator, an advisor, and an evaluator.
  • Enhanced Translation Quality: Achieves competitive performance in English-to-Chinese translation, showing improvements over its base model and other benchmarks on metrics like GRF (87.19), CometKiwi (72.11), and BLEU (36.46).

Good For

  • Research in Reasoning-based MT: Ideal for researchers and developers interested in exploring and advancing long thought reasoning in machine translation.
  • Translating Nuanced Text: Particularly effective for source texts requiring deeper semantic understanding and complex linguistic transfer, such as literary content.
  • Transparent Translation Processes: Provides a unique output format that includes the model's reasoning process, which can be valuable for debugging, analysis, and understanding translation decisions.

Finetunes
1 models