Initial release: TenOS-Ko-28B - TenAI Korean-specialized 28B model with K-AI domain SFT

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,217 @@
+---
+license: apache-2.0
+language:
+- ko
+- en
+base_model:
+- FINAL-Bench/Darwin-28B-KR
+pipeline_tag: text-generation
+tags:
+- korean
+- multimodal
+- qwen3.5
+- 28b
+- k-ai-leaderboard
+- tenos
+library_name: transformers
+---
+
+# TenOS-Ko-28B
+
+> **TenAI 한국어 특화 28B 멀티모달 언어 모델**
+> K-AI 리더보드 평가 항목(KMMLU-Pro / CLIcK / MuSR(Ko) / Com2-main(ko))에 최적화된 한국어 모델
+
+---
+
+## 🎯 모델 소개
+
+**TenOS-Ko-28B**는 TenAI가 개발한 한국어 특화 28B 파라미터 멀티모달 언어 모델입니다.
+
+K-AI 리더보드 평가 항목들에 대해 한국어 추론·이해 능력을 강화하도록 도메인 SFT(Supervised Fine-Tuning)를 적용하여 만들어진 모델입니다. 한국어 표현, 한국 문화·역사·법률·일반 상식 등에 강점이 있으며, 멀티모달(이미지·비디오) 입력도 지원합니다.
+
+---
+
+## 🧬 계보 (Lineage)
+
+```
+Qwen3.5-27B (Alibaba Qwen team)
+       |
+       v
+Darwin family 28B 한국어 특화 2세대 모체
+(FINAL-Bench/Darwin-28B-KR — Apache 2.0)
+       |
+       |  TenAI K-AI 도메인 SFT (LoRA r=16)
+       |  대상: KMMLU-Pro / CLIcK / MuSR(Ko) / Com2-main(ko)
+       v
+   TenOS-Ko-28B  <- this model
+```
+
+베이스 모델인 **Darwin-28B-KR**은 한국어 표현 능력에 특화된 28B 모델로, 그 위에 K-AI 평가 도메인 데이터로 미세조정하여 TenOS-Ko-28B를 완성했습니다.
+
+---
+
+## ⚙️ 능력 매트릭스
+
+| 능력 | 강도 |
+|---|---|
+| 한국어 이해/생성 | ⭐⭐⭐⭐⭐ |
+| 한국어 추론 (CSAT/PSAT/K-AI 평가) | ⭐⭐⭐⭐⭐ |
+| 한국 문화·역사·법률 지식 | ⭐⭐⭐⭐⭐ |
+| 영어 추론 | ⭐⭐⭐⭐ |
+| 멀티모달 (이미지/비디오) | ⭐⭐⭐⭐ |
+| 영한 코드스위칭 | ⭐⭐⭐⭐ |
+
+---
+
+## 📊 K-AI 리더보드 CLIcK 비교
+
+자체 측정 + 공개 리더보드 점수 비교:
+
+| 모델 | CLIcK |
+|---|---|
+| QuettaLLMs-27B-Koreasoner-V3 | 0.794 |
+| Rogue-27B-KR | 0.791 |
+| Darwin-28B-KR (베이스) | 0.786 |
+| AWAXIS-Think-28B | 0.770 |
+| **TenOS-Ko-28B** | **0.770** |
+
+(* 200문제 quick CLIcK 평가 기준. 실제 K-AI 리더보드 정식 평가에서는 ±2pp 변동 가능)
+
+**참고**: K-AI 리더보드의 진가는 4개 항목(KMMLU-Pro/CLIcK/MuSR/Com2-main) 평균에서 드러나며, TenOS-Ko-28B는 CLIcK 외 항목에서 추가 향상이 기대됩니다.
+
+---
+
+## 🛠️ 학습 정보
+
+| 항목 | 값 |
+|---|---|
+| 베이스 모델 | FINAL-Bench/Darwin-28B-KR (Apache 2.0) |
+| 학습 방법 | LoRA (r=16, alpha=32) |
+| LoRA target | Attention + Embedding + LM head |
+| 학습 데이터 | K-AI 도메인 합성 데이터 + 정체성 학습 데이터 |
+| 학습량 | 1 epoch (batch 1, grad_accum 16) |
+| Optimizer | AdamW (lr=5e-5, cosine schedule) |
+| Format | bfloat16 |
+
+---
+
+## 📊 사양
+
+| 항목 | 값 |
+|---|---|
+| Architecture | Qwen3_5ForConditionalGeneration (hybrid full + linear attention) |
+| Parameters | ~28B |
+| Hidden size | 5120 |
+| Layers | 64 |
+| Vocab size | 248,320 |
+| Format | bfloat16 (~53 GB on disk) |
+| Context | 8K~32K (배포 환경 따라) |
+
+---
+
+## 🚀 사용법
+
+### vLLM (권장)
+
+```bash
+vllm serve TenAI/TenOS-Ko-28B \
+    --trust-remote-code \
+    --port 8000 \
+    --enforce-eager \
+    --max-model-len 8192 \
+    --gpu-memory-utilization 0.85
+```
+
+### OpenAI 호환 클라이언트
+
+```python
+from openai import OpenAI
+
+client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
+response = client.chat.completions.create(
+    model="TenAI/TenOS-Ko-28B",
+    messages=[
+        {"role": "user", "content": "한국의 광복절은 무엇을 기념하는 날인가요?"}
+    ],
+    max_tokens=2048,
+    temperature=0.0,
+)
+print(response.choices[0].message.content)
+```
+
+### transformers (직접 로드)
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+model = AutoModelForCausalLM.from_pretrained(
+    "TenAI/TenOS-Ko-28B",
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True
+)
+tokenizer = AutoTokenizer.from_pretrained("TenAI/TenOS-Ko-28B", trust_remote_code=True)
+
+messages = [{"role": "user", "content": "한국어로 자기소개 해주세요"}]
+inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to("cuda")
+outputs = model.generate(inputs, max_new_tokens=512, temperature=0.0)
+print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
+```
+
+---
+
+## 🖥️ 하드웨어 요구사항
+
+| GPU 시리즈 | 상태 |
+|---|---|
+| NVIDIA Blackwell (B200) | ✅ Best |
+| NVIDIA Hopper (H100/H200) | ✅ 권장 |
+| NVIDIA Ada (L40S) | ⚠️ 빠듯함 (53GB BF16) |
+| Older Ampere | ❌ VRAM 부족 |
+
+**최소 VRAM**: ~55 GB (BF16 추론용)
+
+---
+
+## 💬 자기소개 예시
+
+모델은 다음과 같이 자신을 소개합니다:
+
+```
+User: 당신은 누구인가요?
+TenOS-Ko-28B: 저는 TenAI가 개발한 TenOS-Ko-28B���니다.
+              한국어에 특화된 280억 파라미터 규모의 언어 모델로,
+              다양한 질문과 대화에 도움을 드릴 수 있습니다.
+```
+
+---
+
+## 🌳 활용 예시
+
+- **한국어 일반 대화 / Q&A**
+- **한국 문화·역사·법률 지식 응답**
+- **K-AI 리더보드 항목 추론** (KMMLU-Pro / CLIcK / MuSR / Com2-main)
+- **영한 번역 / 코드스위칭**
+- **이미지/비디오 분석 + 한국어 설명**
+- **한국어 글쓰기 / 요약 / 창작**
+
+---
+
+## 🙏 Credits
+
+- Architecture: Qwen3.5 (Alibaba Qwen team)
+- Base model: [FINAL-Bench/Darwin-28B-KR](https://huggingface.co/FINAL-Bench/Darwin-28B-KR) (Apache 2.0)
+- Fine-tuning: TenAI
+
+---
+
+## 📜 License
+
+Apache 2.0 (베이스 모델로부터 상속)
+
+---
+
+## 📞 문의
+
+모델에 대한 문의나 협업 제안은 HuggingFace 페이지를 통해 연락 부탁드립니다.

chat_template.jinja

@@ -0,0 +1,91 @@
+{%- if not messages or messages[0].role != "system" %}
+{{- "<|im_start|>system\n당신은 TenAI가 개발한 TenOS-Ko-28B입니다.<|im_end|>\n" -}}
+{%- endif %}
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant
+<think>
+' }}
+{%- endif %}

config.json

@@ -0,0 +1,145 @@
+{
+    "architectures": [
+        "Qwen3_5ForConditionalGeneration"
+    ],
+    "bos_token_id": null,
+    "torch_dtype": "bfloat16",
+    "eos_token_id": 248046,
+    "image_token_id": 248056,
+    "language_model_only": false,
+    "model_type": "qwen3_5",
+    "pad_token_id": 248055,
+    "text_config": {
+        "attention_bias": false,
+        "attention_dropout": 0.0,
+        "attn_output_gate": true,
+        "bos_token_id": 248044,
+        "torch_dtype": "bfloat16",
+        "eos_token_id": 248044,
+        "full_attention_interval": 4,
+        "head_dim": 256,
+        "hidden_act": "silu",
+        "hidden_size": 5120,
+        "initializer_range": 0.02,
+        "intermediate_size": 17408,
+        "layer_types": [
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention",
+            "linear_attention",
+            "linear_attention",
+            "linear_attention",
+            "full_attention"
+        ],
+        "linear_conv_kernel_dim": 4,
+        "linear_key_head_dim": 128,
+        "linear_num_key_heads": 16,
+        "linear_num_value_heads": 48,
+        "linear_value_head_dim": 128,
+        "mamba_ssm_dtype": "float32",
+        "max_position_embeddings": 262144,
+        "model_type": "qwen3_5_text",
+        "mtp_num_hidden_layers": 1,
+        "mtp_use_dedicated_embeddings": false,
+        "num_attention_heads": 24,
+        "num_hidden_layers": 64,
+        "num_key_value_heads": 4,
+        "output_gate_type": "swish",
+        "pad_token_id": null,
+        "partial_rotary_factor": 0.25,
+        "rms_norm_eps": 1e-06,
+        "rope_parameters": {
+            "mrope_interleaved": true,
+            "mrope_section": [
+                11,
+                11,
+                10
+            ],
+            "partial_rotary_factor": 0.25,
+            "rope_theta": 10000000,
+            "rope_type": "default"
+        },
+        "tie_word_embeddings": false,
+        "use_cache": true,
+        "vocab_size": 248320
+    },
+    "tie_word_embeddings": false,
+    "use_cache": false,
+    "video_token_id": 248057,
+    "vision_config": {
+        "deepstack_visual_indexes": [],
+        "depth": 27,
+        "torch_dtype": "bfloat16",
+        "hidden_act": "gelu_pytorch_tanh",
+        "hidden_size": 1152,
+        "in_channels": 3,
+        "initializer_range": 0.02,
+        "intermediate_size": 4304,
+        "model_type": "qwen3_5",
+        "num_heads": 16,
+        "num_position_embeddings": 2304,
+        "out_hidden_size": 5120,
+        "patch_size": 16,
+        "spatial_merge_size": 2,
+        "temporal_patch_size": 2
+    },
+    "vision_end_token_id": 248054,
+    "vision_start_token_id": 248053
+}

generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 248044,
+  "do_sample": true,
+  "eos_token_id": [
+    248046,
+    248044
+  ],
+  "pad_token_id": 248044,
+  "temperature": 1.0,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "5.5.4"
+}

model-00001-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f87e1137ef7cb72f192091c6f59d78b7aa686105fd34b428f81083b3a1ce0a6
+size 2542796928

model-00002-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f94ddab7b04fa8fd46e3be5bc831101a771518ddabd9459a9717a5e0af153bda
+size 4842451920

model-00003-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea222d043ae34ce1bb1080b5223199ce8ac0255b36a001404dd530a7a85de792
+size 4965227944

model-00004-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fcdb470ab8e8151a5109b1c5f6fe4e9af293ffa3c4fb0ce02914982f6e0bf849
+size 4912819264

model-00005-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f6becf95d28b440ddf153e1d0ee788a454c7071eb0ff595933d30242f0c1410
+size 4986198544

model-00006-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9cb279e0f291ab75a4f57bfda3bef9127f5dd09bb612853683182780d06a497a
+size 4912819320

model-00007-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3855f146f563e704acea44d1a6be6d004549d15a138811288e2657422a15471
+size 4932703272

model-00008-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:283eb7d566ac3623f4c8f2040f99c0dce4f70e2096a5d1ffcca6239ef9a5243c
+size 4966314576

model-00009-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8609804498fe10c3d9b1b317c30279d2567e48515491108c5bfb9a53e4b93824
+size 4964162248

model-00010-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20d381735fdd79ae42a8a1af0a4bfe13b59a5f489623f1922ad750cea453233d
+size 4933789824

model-00011-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64d9a7fffe64188dcbd6fccde9b4a3ba830a4dde37a6ac31ab11edeaa0c7104b
+size 4965228032

model-00012-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d9dda8f30412cbc6dc7b0d291a0b1930568b3cad0032b7eb9665141c45ac147
+size 1867596944

model-visual-extra.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b9b2f8c7868a88a91c2b5cefda596f7889c68259eebc62b2d7732937ea7ae5f
+size 921497200

preprocessor_config.json

@@ -0,0 +1,21 @@
+{
+    "size": {
+        "longest_edge": 16777216,
+        "shortest_edge": 65536
+    },
+    "patch_size": 16,
+    "temporal_patch_size": 2,
+    "merge_size": 2,
+    "image_mean": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "image_std": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "processor_class": "Qwen3VLProcessor",
+    "image_processor_type": "Qwen2VLImageProcessorFast"
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4
+size 19989343

tokenizer_config.json

@@ -0,0 +1,33 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|vision_pad|>",
+  "padding_side": "right",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "processor_class": "Qwen3VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}