Qwen3.5-4B-Base
A base version large language model in the Qwen series with 4 billion parameters, released as open source and commonly used for downstream fine-tuning and domain adaptation research.
Timeline (last 90 days)
The developer expanded the vocabulary of the Qwen3.5-4B-Base model, added keywords missing from languages such as Python, C++, Golang, and Rust, and fine-tuned the model with the extended vocabulary on the SWE dataset.
In the experiment, the original base model (using its own tokenizer) was fine-tuned with the same recipe as a control group; both models used the same training data and pipeline, and the only variable was whether the tokenizer had undergone programming-language-aware extension.
The model with the extended vocabulary improved its pass rate on HumanEval from 45.7% to 52.4%.
On MBPP (Evalplus version), the score of the model with the extended vocabulary dropped slightly from 65.3% to 64.6%.
All Facts (4)
In the experiment, the original base model (using its own tokenizer) was fine-tuned with the same recipe as a control group; both models used the same training data and pipeline, and the only variable was whether the tokenizer had undergone programming-language-aware extension.
50%UnverifiedThe model with the extended vocabulary improved its pass rate on HumanEval from 45.7% to 52.4%.
50%UnverifiedOn MBPP (Evalplus version), the score of the model with the extended vocabulary dropped slightly from 65.3% to 64.6%.
50%UnverifiedThe developer expanded the vocabulary of the Qwen3.5-4B-Base model, added keywords missing from languages such as Python, C++, Golang, and Rust, and fine-tuned the model with the extended vocabulary on the SWE dataset.
50%