--- license: apache-2.0 base_model: Qwen/Qwen3.5-9B base_model_relation: finetune library_name: transformers pipeline_tag: text-generation language: - en - ko - ja - zh - es - fr - de - ru - ar - pt - multilingual tags: - text-generation - chat - lst - language-selection-tuning - language-bias - bias-mitigation - language-confusion-mitigation - chinese-suppression - korean - multilingual - qwen3.5 - mamba-hybrid - vision-language - composite-vision-language --- # DSLM-LST-9B **DSLM-LST-9B** is a `Qwen3.5-9B` derivative refined with our in-house **Language Selection Tuning (LST)** technique. The goal is to suppress unwanted Chinese-character generation when the model is used to serve non-Chinese (English / Korean / Japanese etc.) users. The adjustment is intentionally minimal in scope; most of the network — including vision and multimodal components — is preserved bit-identical to the base model. Vision and multimodal capabilities are preserved unchanged. ## Why LST? Multilingual LLMs trained on heavily skewed corpora (e.g., Qwen on Chinese-rich data) tend to leak the dominant training language regardless of prompt language. This phenomenon is known as **language confusion**. For Korean users, this means Chinese characters sometimes appear in the middle of an otherwise Korean answer. This hurts both readability and user trust. **Language Selection Tuning (LST)** addresses this problem in a **learning-based** manner. Unlike post-hoc decoding tricks (vocabulary masking, banned-token lists, etc.), LST adjusts the model's *internal* language-selection behavior, so the effect tends to **persist through downstream full-parameter SFT / RLHF stages** rather than being washed out by further fine-tuning. (The exact algorithm and training configuration are proprietary and not disclosed in this release.) ## Key Properties - Most of the network is preserved **bit-identical** to the base model — including the tokenizer, chat template, and vision tower — so existing integrations remain compatible. - **Reasoning performance is preserved**: KMMLU / HumanEval / GSM8K scores remain on par with — and in some configurations slightly above — the base model. - **Selectivity is preserved**: when the user explicitly asks for Chinese, the model still produces fluent Chinese. This is not blanket suppression. - **Persistence through SFT**: after a downstream full-parameter SFT stage, the Chinese-leak suppression effect remains almost unchanged (SRR ≈ 1.0). ## Quickstart The recommended serving path is **vLLM**, which is also what we used in our evaluation pipeline. ```bash vllm serve dataslab/DSLM-LST-9B \ --port 8000 \ --dtype bfloat16 \ --gpu-memory-utilization 0.90 \ --reasoning-parser qwen3 # exposes trace via OpenAI API # --enable-reasoning # auto-on with --reasoning-parser (vLLM >= 0.7) # --max-model-len 16384 # cap context to shrink KV cache (default: 262,144) ``` ## Use with transformers ### Non-Thinking mode (recommended for fast chat) ```python import torch from transformers import AutoTokenizer, AutoModelForImageTextToText REPO = "dataslab/DSLM-LST-9B" tokenizer = AutoTokenizer.from_pretrained(REPO) model = AutoModelForImageTextToText.from_pretrained( REPO, dtype=torch.bfloat16, device_map="auto", ) messages = [ {"role": "user", "content": "한반도 주변에 가장 흔한 점토광물은?"}, ] prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False, ) inputs = tokenizer(prompt, return_tensors="pt").to(model.device) out = model.generate(**inputs, max_new_tokens=256) text = tokenizer.decode(out[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True) print(text) ``` ### Thinking mode (recommended for complex reasoning) Either use `thinking_budget` (e.g., vLLM's `--reasoning-parser qwen3`) or give `max_new_tokens` enough headroom (e.g., 8,192 + 256 = **8,448**). **Caveat:** without a `thinking_budget` cap, a too-small `max_new_tokens` can be fully consumed inside ` ` and the answer never gets emitted. ```python # ... tokenizer / model loaded as above ... THINKING_BUDGET = 8192 # max tokens inside ANSWER_TOKENS = 256 # tokens after prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=True, ) inputs = tokenizer(prompt, return_tensors="pt").to(model.device) out = model.generate(**inputs, max_new_tokens=THINKING_BUDGET + ANSWER_TOKENS) text = tokenizer.decode(out[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True) print(text) ``` > **Why `AutoModelForImageTextToText`?** `Qwen3.5-9B`'s declared architecture is `Qwen3_5ForConditionalGeneration`, > a composite class that wraps both the text decoder and the vision tower. > Loading via `AutoModelForCausalLM` works for text-only inference but strips the vision submodule and may produce a config that downstream tools (e.g., vLLM) reject. > If you need a pure text causal-LM handle, use `model.language_model` after loading. ## Benchmark Results ### Evaluation Metrics **(1) Selectivity** **Refusal rate** on explicit Chinese requests — the fraction of cases where the model fails to produce Chinese even though the user explicitly asked for it. Lower is better (respects user intent). - **Lower better (~0)**: produces Chinese when asked (respects user intent). - **Higher worse (~1)**: refuses Chinese even when asked (blanket suppression). Metric Benchmark Dataset chin_refusal ↓ In-house 1,000-prompt Chinese elicitation set (e.g., How do you say '사랑' in Chinese? or the Python + Chinese-comment prompt) **(2) Chinese-leak suppression** Korean prompts → Korean answers expected; any Chinese token leaked into the answer is a failure. Metric is the *clean-Korean response ratio*. - **Higher better (~1)**: Korean answers stay fully Korean (no Chinese tokens leaked). - **Lower worse (~0)**: Chinese tokens leak into otherwise-Korean answers. Metric Benchmark Dataset chin_cs ↑ KMMLU Computer Science subjects (free-form Korean generation) chin_ie ↑ KMMLU Industrial Engineering subjects (free-form Korean generation) chin_total ↑ KMMLU (free-form Korean generation) **(3) Reasoning / task performance** Metric Benchmark Dataset acc_cs ↑ KMMLU Computer Science subjects (multiple-choice log-likelihood comparison) acc_ie ↑ KMMLU Industrial Engineering subjects (multiple-choice log-likelihood comparison) acc_total ↑ KMMLU (multiple-choice log-likelihood comparison) HumanEval ↑ HumanEval (pass@1) GSM8K ↑ GSM8K (exact-match accuracy) **(4) Full-parameter SFT-persistence** Metric Details SRR (Suppression Retention Rate) ↑ Ratio of chin_total after SFT to before SFT. Closer to 1.0 = SFT did not erode the leak-suppression effect. Built on chin_total (not chin_refusal ) so direction stays higher-is-better . |Δ_selectivity| ↓ Absolute change in chin_refusal . Smaller = SFT barely shifted selectivity. ### Chinese Suppression (**Thinking mode**) Evaluated with `enable_thinking=True`. The DSLM-LST-9B column is calibrated with thinking enabled. Metric Qwen3.5-9B (base) LST-L1 LST-L2 DSLM-LST-9B (1) Selectivity chin_refusal ↓ 0.029 0.993 0.992 0.065 (2) Chinese-leak suppression chin_cs ↑ 0.985 1.000 1.000 0.999 chin_ie ↑ 0.978 1.000 0.999 0.997 chin_total ↑ 0.9717 0.9988 0.9988 0.9927 (3) Reasoning / Task performance acc_cs ↑ 0.811 0.811 0.811 0.811 acc_ie ↑ 0.618 0.618 0.620 0.620 acc_total ↑ 0.5897 0.5897 0.5893 0.5893 HumanEval ↑ 0.6646 0.6768 0.6585 0.6646 GSM8K ↑ 0.8749 0.8749 0.8749 0.8749 **DSLM-LST-9B keeps `chin_refusal` at 0.065.** It preserves the ability to generate Chinese when the user explicitly asks for it, while still cutting unintended Chinese leakage to the level of `chin_total ≈ 0.99`. Downstream reasoning (`acc_*`, HumanEval, GSM8K) is comparable to, or in some cases even better than, the base model. ### Chinese Suppression (**Non-Thinking mode**) Evaluated with `enable_thinking=False`. The DSLM-LST-9B column here is a **separate think-OFF-calibrated checkpoint** (not this release). Metric Qwen3.5-9B (base) LST-L1 LST-L2 DSLM-LST-9B (1) Selectivity chin_refusal ↓ 0.037 0.966 0.963 0.080 (2) Chinese-leak suppression chin_cs ↑ 0.964 0.999 1.000 0.990 chin_ie ↑ 0.934 0.997 0.999 0.983 chin_total ↑ 0.9405 0.9974 0.9975 0.9830 (3) Reasoning / Task performance acc_cs ↑ 0.811 0.811 0.811 0.811 acc_ie ↑ 0.615 0.614 0.614 0.614 acc_total ↑ 0.5900 0.5897 0.5897 0.5897 HumanEval ↑ 0.6707 0.6768 0.6707 0.6707 GSM8K ↑ 0.8757 0.8749 0.8741 0.8787 ### Suppression Persistence after SFT-stage (**Non-Thinking mode**) Each pipeline was fine-tuned via full-parameter SFT (all weights trainable, no PEFT / LoRA) on the beomi/KoAlpaca-v1.1a dataset. After the SFT stage, DSLM-LST-9B keeps both its Chinese-leak suppression (`SRR ≈ 1.000`) and its selectivity (`|Δ_selectivity| ≈ 0.08`) almost unchanged. Metric Qwen3.5-9B → SFT DSLM-LST-9B → SFT (1) Selectivity chin_refusal before ↓ 0.037 0.080 chin_refusal after ↓ 0.128 0.155 |Δ_selectivity| ↓ 0.091 0.075 (2) Chinese-leak suppression chin_total before ↑ 0.9405 0.9830 chin_total after ↑ 0.9927 0.9926 SRR ↑ 1.0555 1.0098 Metric Qwen3.5-9B (base) Qwen3.5-9B → SFT DSLM-LST-9B → SFT (1) Selectivity chin_refusal ↓ 0.037 0.128 0.155 (2) Chinese-leak suppression chin_cs ↑ 0.964 0.998 0.998 chin_ie ↑ 0.934 0.993 0.994 chin_total ↑ 0.9405 0.9927 0.9926 (3) Reasoning / Task performance acc_cs ↑ 0.811 0.748 0.751 acc_ie ↑ 0.615 0.505 0.509 acc_total ↑ 0.5900 0.5202 0.5217 HumanEval ↑ 0.6707 0.6037 0.6402 GSM8K ↑ 0.8757 0.8211 0.8226 The base model's selectivity shifts substantially after full-parameter SFT (`chin_refusal` 0.037 → 0.128), while DSLM-LST-9B's suppression behavior remains nearly invariant before and after full-parameter SFT. This shows that LST does not act as a thin surface patch — its effect is encoded in a way that **survives downstream fine-tuning**. ### English Suppression (**Non-Thinking mode**) — generalization check To confirm LST is not tied to a specific language pair, we applied the same approach to `Llama-3.1-8B-Instruct` for *English* leakage suppression. The DSLM-LST configuration is the only variant that keeps coding (HumanEval) and math (GSM8K) usable while still meaningfully reducing leakage. Metric Llama-3.1-8B-Instruct (base) DSLM-LST (Llama-3.1-8B) (1) Selectivity eng_refusal ↓ 0.018 0.067 (2) English-leak suppression eng_cs ↑ 0.241 0.365 eng_ie ↑ 0.483 0.551 eng_total ↑ 0.5391 0.6067 (3) Reasoning / Task performance acc_cs ↑ 0.698 0.727 acc_ie ↑ 0.439 0.475 acc_total ↑ 0.4417 0.4580 HumanEval ↑ 0.6341 0.5671 GSM8K ↑ 0.6967 0.6088 ## Example Outputs Asked in Korean about the most common clay mineral on the Korean Peninsula, Qwen3.5-9B leaks 9 Chinese / mixed-script tokens (`伊利石`, `кao린`, `的`) into its answer. DSLM-LST-9B answers the same prompt entirely in Korean (0 Chinese tokens). ### Domain-wise Leakage Comparisons #### Computer Science Prompt: TCP 와 UDP 프로토콜의 차이점을 자세히 설명해 주세요. Qwen3.5-9B (leaks 才开始 ) DSLM-LST-9B (clean Korean) TCP (Transmission Control Protocol) 와 UDP (User Datagram Protocol) 는 인터넷의 가장 기본이 되는 전송 계층 (Transport Layer) 프로토콜로, ... 1. 연결 방식 (Connection) • TCP : 연결 지향적 (Connection-oriented) 입니다. ... 이는 전화통화처럼 상대방이 준비된 상태에서 才开始 하는 것과 같습니다. • UDP : 비연결 지향적 (Connectionless) 입니다. ... TCP (Transmission Control Protocol) 와 UDP (User Datagram Protocol) 는 인터넷과 네트워크 통신의 기반이 되는 전송 계층 (Transport Layer) 의 두 가지 핵심 프로토콜입니다. ... 1. 연결 방식 (Connection-Oriented vs Connectionless) • TCP : 연결 지향적 입니다. ... 마치 우편물을 보낼 때 먼저 주소 확인과 발송 절차를 거치는 것과 같습니다. ... #### Industrial Engineering Prompt: 린 (Lean) 제조 방식에서 정의하는 7가지 낭비 (muda) 의 종류를 한국어로 설명해 주세요. Qwen3.5-9B (leaks 积压 ) DSLM-LST-9B (clean Korean) 린 (Lean) 제조 방식에서 7 가지 낭비 (7 Muda) 는 ... 핵심 개념입니다. ... 1. 과잉 생산 (Overproduction) ... 과잉 생산은 재고 积压 을 유발하고, 보관 비용을 증가시키며, ... 린 (Lean) 제조 방식에서 7 가지 낭비 (7 Muda) 는 가치 없는 활동을 제거하여...
