Qwen3 235B Model Logits for Sequence Length 8192

8192 is the fixed sequence length for the packed logits data. The dataset contains logits from the Qwen3 235B model, prepared for knowledge distillation. It was created by arcee-ai and uploaded to Hugging Face in December 2025.

Use Cases

• Training a student model to mimic the Qwen3 235B teacher's output logits for sequences of 8192 tokens.
• Analyzing the compressed logit representations using the specified 128-dimensional k-means centroids and polynomial terms.
• Benchmarking knowledge distillation techniques on a dataset with delta encoding and a specific logprob compressor dimensionality of 151,936.

Strengths

• Logits are pre-packed for a consistent sequence length of 8192, simplifying data loading.
• Logprob compressor configuration is explicitly defined with parameters like d=151936 and k=128.

Limitations

• The total number of training examples (rows) is unknown, making it difficult to assess scale.
• The specific content and source texts used to generate the logits are not described, limiting interpretability.

Provenance

Source

arcee-ai/Qwen3-235B-Logits-Packed-8192 repository on Hugging Face.

Collection Method

Logits extracted from the Qwen3 235B model, processed and packed for knowledge distillation.

Freshness

Last updated on 2025-12-10.