akshay-sked/qwen3-14b-svamp-full-dpo

TEXT GENERATIONPricing:Input $0.48 / Output $0.96Concurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 7, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The akshay-sked/qwen3-14b-svamp-full-dpo is a 14 billion parameter language model, fine-tuned from Qwen/Qwen3-14B using Direct Preference Optimization (DPO) on the svamp_final_dpo_train dataset. This model demonstrates a reward accuracy of 1.0 on its evaluation set, indicating strong performance in aligning with preferred responses. It is specifically optimized for tasks related to mathematical reasoning and problem-solving, as suggested by its training on the SVAMP dataset.

Loading preview...

Model Overview

This model, akshay-sked/qwen3-14b-svamp-full-dpo, is a 14 billion parameter language model derived from the Qwen/Qwen3-14B architecture. It has been fine-tuned using Direct Preference Optimization (DPO) on the svamp_final_dpo_train dataset, focusing on enhancing its ability to generate preferred responses.

Key Performance Metrics

During its training and evaluation, the model achieved notable results:

  • Reward Accuracy: 1.0, indicating perfect alignment with chosen responses over rejected ones in the evaluation set.
  • Rewards/margins: 13.5956, reflecting a significant preference for chosen responses.
  • Loss: 0.0006, demonstrating effective optimization during training.

Training Details

The model was trained for 1 epoch with a learning rate of 5e-06, using an AdamW_TORCH_FUSED optimizer and a cosine learning rate scheduler. The training involved a total batch size of 8 with gradient accumulation steps of 8.

Intended Use Cases

Given its fine-tuning on the SVAMP dataset and high reward accuracy, this model is particularly well-suited for:

  • Mathematical Word Problem Solving: Excelling in tasks requiring numerical reasoning and accurate problem interpretation.
  • Preference-aligned Text Generation: Generating responses that are highly aligned with specified preferences, as indicated by its DPO training.