AmberYifan/capsdnum-marin-8b-base-math_cap_b1000_s0

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Jul 16, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The AmberYifan/capsdnum-marin-8b-base-math_cap_b1000_s0 model is an 8 billion parameter language model, fine-tuned from marin-community/marin-8b-base. It was specifically trained on the capsd_marin-8b-base-n80000-numina__mix_math_cap_b1000_s0 dataset, suggesting an optimization for mathematical reasoning tasks. This model is intended for applications requiring enhanced numerical and mathematical capabilities within an 8192-token context window.

Loading preview...

Model Overview

The AmberYifan/capsdnum-marin-8b-base-math_cap_b1000_s0 is an 8 billion parameter language model, derived from the marin-community/marin-8b-base architecture. It has been fine-tuned with a specific focus on mathematical capabilities, utilizing the capsd_marin-8b-base-n80000-numina__mix_math_cap_b1000_s0 dataset.

Training Details

The model underwent a fine-tuning process with the following key hyperparameters:

  • Learning Rate: 1e-05
  • Batch Sizes: A train_batch_size of 2 and eval_batch_size of 8, with a total_train_batch_size of 64 and total_eval_batch_size of 32, achieved through gradient accumulation over 8 steps.
  • Optimizer: AdamW with betas=(0.9, 0.999) and epsilon=1e-08.
  • Scheduler: Cosine learning rate scheduler with 0.03 warmup steps.
  • Epochs: Trained for 1 epoch.

Intended Use Cases

Given its specialized training on a mathematical dataset, this model is likely best suited for:

  • Mathematical Reasoning: Tasks involving numerical problem-solving, equation handling, and quantitative analysis.
  • Scientific Applications: Potentially useful in fields requiring precise numerical understanding and generation.

Limitations

As indicated in the original model card, further information regarding specific limitations and broader intended uses is needed for a comprehensive understanding of its scope and performance boundaries.

Finetunes
1 models