DeepSeek Mixed-Mode Reasoning Logits Packed for 16384 Sequence Length

A dataset of teacher model logits for knowledge distillation, packed for a sequence length of 16384. It was created by arcee-ai and uploaded to Hugging Face in December 2025. The logits are derived from the DeepSeek-MixedModeReasoning model and are pre-packed for efficient training.

Use Cases

• Distill reasoning capabilities from the teacher model into a smaller student model using the provided logits as training targets.
• Train models on long-sequence reasoning tasks using the pre-packed data formatted for a 16384-token context.
• Analyze the teacher model's output distribution across a vocabulary of 129280 tokens for specific reasoning examples.
• Experiment with logit compression techniques, as the data uses a legacy compression method with parameters like exact_k=32 and polynomial_degree=0.

Strengths

• Logits are pre-packed for a long sequence length of 16384 tokens, optimizing for training efficiency.
• Logit compression is applied with a defined vocabulary size of 129280 and a compression parameter k=32.

Limitations

• The number of training examples, dataset size, and specific source data composition are not provided.
• The 'legacy_logit_compression' method may be less common or require specific handling compared to standard techniques.

Provenance

Source

arcee-ai/DeepSeek-MixedModeReasoning-Logits-Packed-16384 on Hugging Face.

Collection Method

Logits generated by a teacher model (likely DeepSeek-MixedModeReasoning) for knowledge distillation, subsequently packed.

Freshness

Last updated on the Hugging Face Hub in December 2025.