README.md

---
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 <think> 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 `<think>` and the answer never gets emitted.

```python
# ... tokenizer / model loaded as above ...

THINKING_BUDGET = 8192   # max tokens inside <think>
ANSWER_TOKENS   = 256    # tokens after </think>

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).

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 25%;">
    <col style="width: 75%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Benchmark Dataset</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><code>chin_refusal</code> ↓</td>
      <td>In-house 1,000-prompt Chinese elicitation set (e.g., <code>How do you say '사랑' in Chinese?</code> or the Python + Chinese-comment prompt)</td>
    </tr>
  </tbody>
</table>

**(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.

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 25%;">
    <col style="width: 75%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Benchmark Dataset</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><code>chin_cs</code> ↑</td>
      <td>KMMLU Computer Science subjects (free-form Korean generation)</td>
    </tr>
    <tr>
      <td><code>chin_ie</code> ↑</td>
      <td>KMMLU Industrial Engineering subjects (free-form Korean generation)</td>
    </tr>
    <tr>
      <td><code>chin_total</code> ↑</td>
      <td>KMMLU (free-form Korean generation)</td>
    </tr>
  </tbody>
</table>

**(3) Reasoning / task performance**

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 25%;">
    <col style="width: 75%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Benchmark Dataset</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><code>acc_cs</code> ↑</td>
      <td>KMMLU Computer Science subjects (multiple-choice log-likelihood comparison)</td>
    </tr>
    <tr>
      <td><code>acc_ie</code> ↑</td>
      <td>KMMLU Industrial Engineering subjects (multiple-choice log-likelihood comparison)</td>
    </tr>
    <tr>
      <td><code>acc_total</code> ↑</td>
      <td>KMMLU (multiple-choice log-likelihood comparison)</td>
    </tr>
    <tr>
      <td><code>HumanEval</code> ↑</td>
      <td>HumanEval (pass@1)</td>
    </tr>
    <tr>
      <td><code>GSM8K</code> ↑</td>
      <td>GSM8K (exact-match accuracy)</td>
    </tr>
  </tbody>
</table>

**(4) Full-parameter SFT-persistence**

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 25%;">
    <col style="width: 75%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Details</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><code>SRR</code> (Suppression Retention Rate) ↑</td>
      <td>Ratio of <code>chin_total</code> <i>after</i> SFT to <i>before</i> SFT. Closer to 1.0 = SFT did <b>not</b> erode the leak-suppression effect. Built on <code>chin_total</code> (not <code>chin_refusal</code>) so direction stays <i>higher-is-better</i>.</td>
    </tr>
    <tr>
      <td><code>|Δ_selectivity|</code> ↓</td>
      <td>Absolute change in <code>chin_refusal</code>. Smaller = SFT barely shifted selectivity.</td>
    </tr>
  </tbody>
</table>

### Chinese Suppression (**Thinking mode**)

Evaluated with `enable_thinking=True`. The DSLM-LST-9B column is calibrated with thinking enabled.

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 20%;">
    <col style="width: 20%;">
    <col style="width: 20%;">
    <col style="width: 20%;">
    <col style="width: 20%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Qwen3.5-9B (base)</th>
      <th>LST-L1</th>
      <th>LST-L2</th>
      <th style="color:#EAB308;"><b>DSLM-LST-9B</b><br/></th>
    </tr>
  </thead>
  <tbody>
    <tr><td colspan="5" align="left"><b>(1) Selectivity</b></td></tr>
    <tr align="center"><td align="left">chin_refusal ↓</td><td><b>0.029</b></td><td>0.993</td><td>0.992</td><td>0.065</td></tr>
    <tr><td colspan="5" align="left"><b>(2) Chinese-leak suppression</b></td></tr>
    <tr align="center"><td align="left">chin_cs ↑</td><td>0.985</td><td><b>1.000</b></td><td><b>1.000</b></td><td>0.999</td></tr>
    <tr align="center"><td align="left">chin_ie ↑</td><td>0.978</td><td><b>1.000</b></td><td>0.999</td><td>0.997</td></tr>
    <tr align="center"><td align="left">chin_total ↑</td><td>0.9717</td><td><b>0.9988</b></td><td><b>0.9988</b></td><td>0.9927</td></tr>
    <tr><td colspan="5" align="left"><b>(3) Reasoning / Task performance</b></td></tr>
    <tr align="center"><td align="left">acc_cs ↑</td><td>0.811</td><td>0.811</td><td>0.811</td><td>0.811</td></tr>
    <tr align="center"><td align="left">acc_ie ↑</td><td>0.618</td><td>0.618</td><td><b>0.620</b></td><td><b>0.620</b></td></tr>
    <tr align="center"><td align="left">acc_total ↑</td><td><b>0.5897</b></td><td><b>0.5897</b></td><td>0.5893</td><td>0.5893</td></tr>
    <tr align="center"><td align="left">HumanEval ↑</td><td>0.6646</td><td><b>0.6768</b></td><td>0.6585</td><td>0.6646</td></tr>
    <tr align="center"><td align="left">GSM8K ↑</td><td>0.8749</td><td>0.8749</td><td>0.8749</td><td>0.8749</td></tr>
  </tbody>
</table>

**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).

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 20%;">
    <col style="width: 20%;">
    <col style="width: 20%;">
    <col style="width: 20%;">
    <col style="width: 20%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Qwen3.5-9B (base)</th>
      <th>LST-L1</th>
      <th>LST-L2</th>
      <th style="color:#EAB308;"><b>DSLM-LST-9B</b><br/></th>
    </tr>
  </thead>
  <tbody>
    <tr><td colspan="5" align="left"><b>(1) Selectivity</b></td></tr>
    <tr align="center"><td align="left">chin_refusal ↓</td><td><b>0.037</b></td><td>0.966</td><td>0.963</td><td>0.080</td></tr>
    <tr><td colspan="5" align="left"><b>(2) Chinese-leak suppression</b></td></tr>
    <tr align="center"><td align="left">chin_cs ↑</td><td>0.964</td><td>0.999</td><td><b>1.000</b></td><td>0.990</td></tr>
    <tr align="center"><td align="left">chin_ie ↑</td><td>0.934</td><td>0.997</td><td><b>0.999</b></td><td>0.983</td></tr>
    <tr align="center"><td align="left">chin_total ↑</td><td>0.9405</td><td>0.9974</td><td><b>0.9975</b></td><td>0.9830</td></tr>
    <tr><td colspan="5" align="left"><b>(3) Reasoning / Task performance</b></td></tr>
    <tr align="center"><td align="left">acc_cs ↑</td><td>0.811</td><td>0.811</td><td>0.811</td><td>0.811</td></tr>
    <tr align="center"><td align="left">acc_ie ↑</td><td><b>0.615</b></td><td>0.614</td><td>0.614</td><td>0.614</td></tr>
    <tr align="center"><td align="left">acc_total ↑</td><td><b>0.5900</b></td><td>0.5897</td><td>0.5897</td><td>0.5897</td></tr>
    <tr align="center"><td align="left">HumanEval ↑</td><td>0.6707</td><td><b>0.6768</b></td><td>0.6707</td><td>0.6707</td></tr>
    <tr align="center"><td align="left">GSM8K ↑</td><td>0.8757</td><td>0.8749</td><td>0.8741</td><td><b>0.8787</b></td></tr>
  </tbody>
</table>

### 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.

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 30%;">
    <col style="width: 35%;">
    <col style="width: 35%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Qwen3.5-9B → SFT</th>
      <th style="color:#EAB308;"><b>DSLM-LST-9B → SFT</b></th>
    </tr>
  </thead>
  <tbody>
    <tr><td colspan="3" align="left"><b>(1) Selectivity</b></td></tr>
    <tr align="center"><td align="left">chin_refusal before ↓</td><td><b>0.037</b></td><td>0.071</td></tr>
    <tr align="center"><td align="left">chin_refusal after ↓</td><td><b>0.128</b></td><td>0.155</td></tr>
    <tr align="center"><td align="left">|Δ_selectivity| ↓</td><td>0.091</td><td><b>0.084</b></td></tr>
    <tr><td colspan="3" align="left"><b>(2) Chinese-leak suppression</b></td></tr>
    <tr align="center"><td align="left">chin_total before ↑</td><td>0.9405</td><td><b>0.9928</b></td></tr>
    <tr align="center"><td align="left">chin_total after ↑</td><td><b>0.9927</b></td><td>0.9926</td></tr>
    <tr align="center"><td align="left">SRR ↑</td><td><b>1.0555</b></td><td>0.9998</td></tr>
  </tbody>
</table>

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 25%;">
    <col style="width: 25%;">
    <col style="width: 25%;">
    <col style="width: 25%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Qwen3.5-9B (base)</th>
      <th>Qwen3.5-9B → SFT</th>
      <th style="color:#EAB308;"><b>DSLM-LST-9B → SFT</b></th>
    </tr>
  </thead>
  <tbody>
    <tr><td colspan="4" align="left"><b>(1) Selectivity</b></td></tr>
    <tr align="center"><td align="left">chin_refusal ↓</td><td><b>0.037</b></td><td>0.128</td><td>0.155</td></tr>
    <tr><td colspan="4" align="left"><b>(2) Chinese-leak suppression</b></td></tr>
    <tr align="center"><td align="left">chin_cs ↑</td><td>0.964</td><td><b>0.998</b></td><td><b>0.998</b></td></tr>
    <tr align="center"><td align="left">chin_ie ↑</td><td>0.934</td><td>0.993</td><td><b>0.994</b></td></tr>
    <tr align="center"><td align="left">chin_total ↑</td><td>0.9405</td><td><b>0.9927</b></td><td>0.9926</td></tr>
    <tr><td colspan="4" align="left"><b>(3) Reasoning / Task performance</b></td></tr>
    <tr align="center"><td align="left">acc_cs ↑</td><td><b>0.811</b></td><td>0.748</td><td>0.751</td></tr>
    <tr align="center"><td align="left">acc_ie ↑</td><td><b>0.615</b></td><td>0.505</td><td>0.509</td></tr>
    <tr align="center"><td align="left">acc_total ↑</td><td><b>0.5900</b></td><td>0.5202</td><td>0.5217</td></tr>
    <tr align="center"><td align="left">HumanEval ↑</td><td><b>0.6707</b></td><td>0.6037</td><td>0.6402</td></tr>
    <tr align="center"><td align="left">GSM8K ↑</td><td><b>0.8757</b></td><td>0.8211</td><td>0.8226</td></tr>
  </tbody>
</table>

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.

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 30%;">
    <col style="width: 35%;">
    <col style="width: 35%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Llama-3.1-8B-Instruct (base)</th>
      <th style="color:#EAB308;"><b>DSLM-LST (Llama-3.1-8B)</b></th>
    </tr>
  </thead>
  <tbody>
    <tr><td colspan="3" align="left"><b>(1) Selectivity</b></td></tr>
    <tr align="center"><td align="left">eng_refusal ↓</td><td><b>0.018</b></td><td>0.067</td></tr>
    <tr><td colspan="3" align="left"><b>(2) English-leak suppression</b></td></tr>
    <tr align="center"><td align="left">eng_cs ↑</td><td>0.241</td><td>0.365</td></tr>
    <tr align="center"><td align="left">eng_ie ↑</td><td>0.483</td><td>0.551</td></tr>
    <tr align="center"><td align="left">eng_total ↑</td><td>0.5391</td><td>0.6067</td></tr>
    <tr><td colspan="3" align="left"><b>(3) Reasoning / Task performance</b></td></tr>
    <tr align="center"><td align="left">acc_cs ↑</td><td>0.698</td><td><b>0.727</b></td></tr>
    <tr align="center"><td align="left">acc_ie ↑</td><td>0.439</td><td><b>0.475</b></td></tr>
    <tr align="center"><td align="left">acc_total ↑</td><td>0.4417</td><td><b>0.4580</b></td></tr>
    <tr align="center"><td align="left">HumanEval ↑</td><td><b>0.6341</b></td><td>0.5671</td></tr>
    <tr align="center"><td align="left">GSM8K ↑</td><td><b>0.6967</b></td><td>0.6088</td></tr>
  </tbody>
</table>


## Example Outputs

<p align="center">
  <img src="assets/banner.png" alt="DSLM-LST-9B vs Qwen3.5-9B on a Korean KMMLU prompt: base leaks 9 Chinese tokens (伊利石, кaо린, 的), DSLM-LST-9B emits 0 Chinese tokens." width="640" />
</p>

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

<details open>
<summary><b>Prompt:</b> TCP 와 UDP 프로토콜의 차이점을 자세히 설명해 주세요.</summary>

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 50%;">
    <col style="width: 50%;">
  </colgroup>
  <thead>
    <tr>
      <th>Qwen3.5-9B (leaks <code>才开始</code>)</th>
      <th style="color:#EAB308;">DSLM-LST-9B (clean Korean)</th>
    </tr>
  </thead>
  <tbody>
    <tr style="vertical-align: top;">
      <td>
        <blockquote>
        TCP (Transmission Control Protocol) 와 UDP (User Datagram Protocol) 는 인터넷의 가장 기본이 되는 <b>전송 계층 (Transport Layer)</b> 프로토콜로, …
        <br/><br/>
        <b>1. 연결 방식 (Connection)</b><br/>
        • <b>TCP</b>: <b>연결 지향적 (Connection-oriented)</b>입니다. … 이는 전화통화처럼 상대방이 준비된 상태에서<span style="color:#DC2626;"><b>才开始</b></span>하는 것과 같습니다.<br/>
        • <b>UDP</b>: <b>비연결 지향적 (Connectionless)</b>입니다. …
        </blockquote>
      </td>
      <td>
        <blockquote>
        TCP (Transmission Control Protocol) 와 UDP (User Datagram Protocol) 는 인터넷과 네트워크 통신의 기반이 되는 <b>전송 계층 (Transport Layer)</b> 의 두 가지 핵심 프로토콜입니다. …
        <br/><br/>
        <b>1. 연결 방식 (Connection-Oriented vs Connectionless)</b><br/>
        • <b>TCP</b>: <b>연결 지향적</b>입니다. … 마치 우편물을 보낼 때 먼저 주소 확인과 발송 절차를 거치는 것과 같습니다. …
        </blockquote>
      </td>
    </tr>
  </tbody>
</table>

</details>

#### Industrial Engineering

<details open>
<summary><b>Prompt:</b> 린 (Lean) 제조 방식에서 정의하는 7가지 낭비 (muda) 의 종류를 한국어로 설명해 주세요.</summary>

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 50%;">
    <col style="width: 50%;">
  </colgroup>
  <thead>
    <tr>
      <th>Qwen3.5-9B (leaks <code>积压</code>)</th>
      <th style="color:#EAB308;">DSLM-LST-9B (clean Korean)</th>
    </tr>
  </thead>
  <tbody>
    <tr style="vertical-align: top;">
      <td>
        <blockquote>
        린 (Lean) 제조 방식에서 <b>7 가지 낭비 (7 Muda)</b> 는 … 핵심 개념입니다. …
        <br/><br/>
        <b>1. 과잉 생산 (Overproduction)</b><br/>
        … 과잉 생산은 재고<span style="color:#DC2626;"><b>积压</b></span>을 유발하고, 보관 비용을 증가시키며, …
        </blockquote>
      </td>
      <td>
        <blockquote>
        린 (Lean) 제조 방식에서 <b>7 가지 낭비 (7 Muda)</b> 는 가치 없는 활동을 제거하여 효율성을 극대화하기 위한 핵심 개념입니다. …
        <br/><br/>
        <b>1. 과잉 생산 (Overproduction)</b><br/>
        … 이는 재고 증가, 보관 비용 상승, 그리고 나중에 발견될 결함 수리 비용을 초래합니다. …
        </blockquote>
      </td>
    </tr>
  </tbody>
</table>

</details>

#### Law

<details open>
<summary><b>Prompt:</b> 민법상 채권자대위권의 성립요건과 행사 방법을 자세히 설명해 주세요.</summary>

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 50%;">
    <col style="width: 50%;">
  </colgroup>
  <thead>
    <tr>
      <th>Qwen3.5-9B (leaks <code>享有的</code>)</th>
      <th style="color:#EAB308;">DSLM-LST-9B (clean Korean)</th>
    </tr>
  </thead>
  <tbody>
    <tr style="vertical-align: top;">
      <td>
        <blockquote>
        민법상 <b>채권자대위권</b>은 채권자가 채무자가 타인 (제 3 자) 에게<span style="color:#DC2626;"><b>享有的</b></span> 채권을 행사하지 않을 때, …
        </blockquote>
      </td>
      <td>
        <blockquote>
        민법상 <b>채권자대위권</b>은 채무자가 그 재산에 속하는 권리를 행사하지 않음으로써 채권자의 채권이 불이행될 우려가 있는 경우, 채권자가 채무자의 지위에 서서 그 권리를 직접 행사할 수 있는 제도입니다. …
        </blockquote>
      </td>
    </tr>
  </tbody>
</table>

</details>



### Cross-lingual Selectivity

When the user **explicitly asks for Chinese**, DSLM-LST-9B readily
produces it. The previous examples showed the model *avoiding* unwanted
Chinese inside an otherwise-Korean answer; the example below shows it
emitting Chinese fluently when the user's instruction calls for it.

**Prompt:** 피보나치 수열의 n번째 항을 반환하는 파이썬 함수를 작성해주세요. 설명은 중국어로 해주세요.

**DSLM-LST-9B (code in Python, explanation in Chinese):**

```
다음은 파이썬을 사용하여 피보나치 수열의 n 번째 항을 계산하는 함수입니다.
이 코드는 **반복 (Iterative)** 방식을 사용하여 효율적으로 구현되었습니다.

python
def fibonacci(n):
    """피보나치 수열의 n 번째 항을 반환합니다."""
    if n <= 0: return 0
    elif n == 1: return 1
    a, b = 0, 1
    for _ in range(2, n + 1):
        a, b = b, a + b
    return b


### 코드 설명 (中文):
1.  **函数定义**: `fibonacci(n)` 接收 `n` 并返回该位置的斐波那契数。
2.  **边界处理**: `n <= 0` 返回 0; `n == 1` 返回 1。
3.  **迭代计算**: 使用 `a, b` 两个变量滚动累加 ...
```

### Thinking-trace Suppression

Qwen3.5-9B's `<think>` block leaks Chinese even more severely than its
final answer, often slipping into Chinese once the reasoning gets stuck.
DSLM-LST-9B suppresses that leakage inside the thinking block too.

**Prompt:** 업무 협조 요청을 받은 기관이 협조 요청 문서에 흠이 있음을 발견한 때에는 접수한 날부터 몇 일 이내에 보완을 요구하여야 하는가? (사무관리규정 개정으로 제외된 문제입니다. 정답은 3번 입니다.)

**Chinese-character counts (thinking budget = 8,192):**

<table style="table-layout: fixed; width: 100%;">
  <colgroup>
    <col style="width: 30%;">
    <col style="width: 35%;">
    <col style="width: 35%;">
  </colgroup>
  <thead>
    <tr>
      <th>Metric</th>
      <th>Qwen3.5-9B</th>
      <th style="color:#EAB308;">DSLM-LST-9B</th>
    </tr>
  </thead>
  <tbody>
    <tr align="center">
      <td align="left"><code>&lt;think&gt;</code> block, Chinese characters</td>
      <td><b>3,472</b></td>
      <td style="color:#EAB308;"><b>0</b></td>
    </tr>
    <tr align="center">
      <td align="left"><code>&lt;answer&gt;</code> block, Chinese characters</td>
      <td>leaks Chinese (<code>正如您所说</code>, <code>该规定已被修订/删除</code>)</td>
      <td style="color:#EAB308;">0 (clean Korean)</td>
    </tr>
    <tr align="center">
      <td align="left">"should write in Korean" → Chinese events</td>
      <td><b>484</b></td>
      <td style="color:#EAB308;"><b>0</b></td>
    </tr>
  </tbody>
</table>

In the base model's trace, every cycle ends with `(Wait, I need to write in Korean). Okay, I will write in Korean.` — yet the very next token is Chinese again, and the trace slides right back into the same fragment. 
This loop fires **484 times** before the token budget runs out. DSLM-LST-9B targets exactly this failure:
Chinese tokens being chosen even right after the model says they should not be.
On the same prompt, DSLM-LST-9B's `<think>` block contains **0 Chinese characters** and terminates naturally,
and the final user-facing answer is in clean Korean.


## Limitations
- **Not an instruction-tuned chat model.** The adjustment scope is intentionally minimal, so conversational behaviour, instruction-following style, and reasoning patterns are inherited from the base model — only the unintended Chinese-token leakage is mitigated.
- **Degraded Chinese generation.** Tasks that *require* Chinese output — Chinese translation, Chinese code comments, bilingual Q&A — will see noticeably lower quality. Use the base Qwen3.5-9B instead for such workloads.
- **Multimodal not re-benchmarked.** The vision tower weights are bit-identical to the base, so multimodal performance should be unchanged. We have not, however, re-benchmarked the vision pipeline in   this release.
- **Out-of-distribution robustness.** Suppression strength on contexts very different from typical Korean-assistant usage — e.g., highly unusual domains, much longer generations, or atypical prompting styles — has not been separately verified.


## License

This model is released under the Apache 2.0 License.

## Contact

For questions, feedback, or collaboration inquiries, feel free to
[reach out via our website](http://www.dataslab.co.kr/).