dataslab/DSLM-LST-9B
dataslab/DSLM-LST-9B is a 9-billion parameter Qwen3.5 derivative developed by dataslab, specifically refined using Language Selection Tuning (LST). This model is designed to suppress unwanted Chinese character generation in non-Chinese outputs, particularly for users of languages like English, Korean, and Japanese, while preserving vision and multimodal capabilities. It maintains reasoning performance comparable to its base model and ensures the Chinese-leak suppression effect persists through downstream fine-tuning stages.
Loading preview...
Overview
dataslab/DSLM-LST-9B is a 9-billion parameter model based on Qwen3.5, developed by dataslab. Its core innovation is Language Selection Tuning (LST), a learning-based technique designed to prevent unintended Chinese character leakage in responses to non-Chinese prompts (e.g., Korean, English, Japanese). Unlike post-hoc decoding methods, LST modifies the model's internal language-selection behavior, ensuring its effect is robust and persists even after further full-parameter fine-tuning (SFT/RLHF).
Key Capabilities & Features
- Chinese-Leak Suppression: Significantly reduces the occurrence of Chinese characters in non-Chinese outputs, improving readability and user trust for multilingual applications.
- Preserved Selectivity: The model retains the ability to generate fluent Chinese when explicitly requested by the user, avoiding blanket suppression.
- Reasoning Performance: Benchmarks on KMMLU, HumanEval, and GSM8K show reasoning and task performance remain on par with, or slightly exceed, the base Qwen3.5-9B model.
- Persistence Through SFT: The Chinese-leak suppression effect is highly stable and largely unaffected by subsequent full-parameter SFT stages, as demonstrated by a Suppression Retention Rate (SRR) close to 1.0.
- Bit-Identical Core: Most of the network, including the tokenizer, chat template, and vision tower, is preserved bit-identical to the base model, ensuring compatibility with existing integrations and retaining multimodal capabilities.
Use Cases
- Multilingual Applications: Ideal for applications serving non-Chinese users where unintended Chinese output is undesirable, particularly for Korean, English, and Japanese language tasks.
- Downstream Fine-tuning: Suitable as a base for further fine-tuning, as its core language selection improvements are designed to persist.
- Complex Reasoning: Supports a "Thinking mode" for complex reasoning tasks, with suppressed Chinese leakage even within internal thought processes.
Limitations
- Not an Instruction-Tuned Chat Model: Inherits conversational behavior and instruction-following style from the base model; LST primarily addresses language leakage.
- Degraded Chinese Generation: Quality for tasks explicitly requiring Chinese output (e.g., translation, Chinese code comments) will be lower than the base Qwen3.5-9B.
- Multimodal Benchmarking: While vision capabilities are preserved, they have not been re-benchmarked in this release.