| --- |
| license: apache-2.0 |
| language: |
| - ko |
| - en |
| pipeline_tag: text-generation |
| library_name: transformers |
| base_model: PoSTMEDIA/Rosetta-7B-Base |
| tags: |
| - rosetta |
| - korean |
| - bilingual |
| - postmedia |
| --- |
| |
| <div align="center"> |
|
|
| # Rosetta-7B-Instruct |
|
|
| [](https://huggingface.co/collections/PoSTMEDIA/rosetta-6a9db30fd1b4585b0c1845e9) |
| [](https://github.com/PoSTMEDIA-AI/vllm) |
| [](./LICENSE) |
|
|
| </div> |
|
|
| ## Introduction |
|
|
| **Rosetta-7B-Instruct** is a 7B-parameter bilingual (Korean-English) instruction-tuned model developed by **PoSTMEDIA**. Built on PoSTMEDIA's **Rosetta** dense decoder-only architecture, it is post-trained from [Rosetta-7B-Base](https://huggingface.co/PoSTMEDIA/Rosetta-7B-Base) through large-scale supervised fine-tuning and preference optimization, with a deliberate focus on the capabilities that matter most in real Korean-language deployments: instruction following, Korean cultural and factual knowledge, and Korean mathematical reasoning. |
|
|
| Rosetta-7B is a three-model family covering the full spectrum from research to production: |
|
|
| | Model | Download | Note | |
| |---|---|---| |
| | Rosetta-7B-Base | [HuggingFace](https://huggingface.co/PoSTMEDIA/Rosetta-7B-Base) | Foundation model (completion-style) | |
| | **Rosetta-7B-Instruct** | [HuggingFace](https://huggingface.co/PoSTMEDIA/Rosetta-7B-Instruct) | Instruction following / chat *(this model)* | |
| | Rosetta-7B-Instruct-NVFP4 | [HuggingFace](https://huggingface.co/PoSTMEDIA/Rosetta-7B-Instruct-NVFP4) | NVFP4 4-bit of Instruct (NVIDIA Blackwell / DGX Spark) | |
| | Rosetta-7B-Think | [HuggingFace](https://huggingface.co/PoSTMEDIA/Rosetta-7B-Think) | Explicit reasoning (`<think>`) | |
| | Rosetta-7B-Think-NVFP4 | [HuggingFace](https://huggingface.co/PoSTMEDIA/Rosetta-7B-Think-NVFP4) | NVFP4 4-bit of Think (NVIDIA Blackwell / DGX Spark) | |
|
|
| ## Highlights |
|
|
| - **Korean math reasoning leader** β HRM8K **60.4**, the top score in our eight-model same-protocol comparison of Korean and global open models |
| - **Top-tier instruction following** β IFEval **80.4**, within the leading group of Korean open models |
| - **Korean-first design** β 161K Korean-extended vocabulary, dedicated Korean continual-pretraining stage, and in-house synthetic Korean data assets |
| - **Strong Korean factuality** β KoSimpleQA **71.7** |
| - **65,536-token context window** with interleaved local-global attention |
| - **Apache-2.0** β unrestricted commercial use |
|
|
| ## Model Summary |
|
|
| <div align="center"> |
| <table><tbody> |
| <tr><td><b>Architecture</b></td><td>Rosetta dense decoder-only Transformer (<code>RosettaForCausalLM</code>)</td></tr> |
| <tr><td><b>Parameters</b></td><td>7B</td></tr> |
| <tr><td><b>Layers</b></td><td>32</td></tr> |
| <tr><td><b>Hidden size</b></td><td>4,096</td></tr> |
| <tr><td><b>Attention heads</b></td><td>32</td></tr> |
| <tr><td><b>Attention pattern</b></td><td>interleaved sliding-window (4,096) + global, 3:1, with QK-normalization</td></tr> |
| <tr><td><b>Context length</b></td><td>65,536</td></tr> |
| <tr><td><b>Vocabulary</b></td><td>161,425 (Korean-extended)</td></tr> |
| <tr><td><b>Post-training</b></td><td>SFT β preference optimization (DPO)</td></tr> |
| <tr><td><b>License</b></td><td>Apache-2.0</td></tr> |
| </tbody></table> |
| </div> |
|
|
| ## Training Overview |
|
|
| The Rosetta-7B family was built through a multi-stage pipeline carried out end-to-end by PoSTMEDIA: |
|
|
| 1. **Pretraining** on trillions of tokens of curated bilingual web, code, and academic text |
| 2. **Staged mid-training** for reasoning-dense data and long-context extension up to 65K |
| 3. **Korean continual pretraining** on curated Korean corpora plus in-house synthetic Korean data assets, with the vocabulary extended to 161K for efficient Korean tokenization |
| 4. **Post-training** β large-scale supervised fine-tuning followed by preference optimization (DPO) targeting instruction following, factuality, and safe refusals |
|
|
| ## Evaluation Results |
|
|
| All models in the table below, **including competitors, were re-evaluated in-house under an identical protocol** (lm-evaluation-harness + vLLM β₯ 0.26, identical prompts, decoding parameters, and generation budgets). Scores are therefore directly comparable within this table, but may differ from numbers reported elsewhere under different setups. |
|
|
| <!-- prettier-ignore --> |
| <table> |
| <tr> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;">Benchmark</th> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;"><b>Rosetta-7B-Instruct</b><br><sup>7B</sup></th> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;">A.X-4.0-Light<br><sup>7B</sup></th> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;">A.X-3.1-Light<br><sup>7B</sup></th> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;">EXAONE-3.5<br><sup>7.8B</sup></th> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;">Midm-2.0-Base<br><sup>11.5B</sup></th> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;">kanana-2<br><sup>3B</sup></th> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;">Llama-3.1<br><sup>8B</sup></th> |
| <th style="background: rgba(128,128,128,0.1); text-align: center;">Ministral-3<br><sup>8B</sup></th> |
| </tr> |
| <tr> |
| <td align="center" colspan="9" style="background: linear-gradient(90deg, rgba(0,82,204,0.30) 0%, rgba(0,163,255,0.30) 50%, rgba(120,86,255,0.30) 100%); font-weight: bold; height: 32px; padding-top: 2px; padding-bottom: 2px;"><i>General & Reasoning</i></td> |
| </tr> |
| <tr> |
| <td>MMLU</td> |
| <td align="center">60.2</td> |
| <td align="center">73.0</td> |
| <td align="center">57.7</td> |
| <td align="center"><b>73.4</b></td> |
| <td align="center">68.0</td> |
| <td align="center">59.5</td> |
| <td align="center">66.6</td> |
| <td align="center">72.4</td> |
| </tr> |
| <tr> |
| <td>GPQA</td> |
| <td align="center">39.4</td> |
| <td align="center">37.4</td> |
| <td align="center">30.8</td> |
| <td align="center">36.9</td> |
| <td align="center">31.8</td> |
| <td align="center">32.3</td> |
| <td align="center">22.7</td> |
| <td align="center"><b>49.5</b></td> |
| </tr> |
| <tr> |
| <td>GSM8K</td> |
| <td align="center">79.8</td> |
| <td align="center">68.1</td> |
| <td align="center">84.0</td> |
| <td align="center"><b>88.9</b></td> |
| <td align="center">80.7</td> |
| <td align="center">70.0</td> |
| <td align="center">84.2</td> |
| <td align="center">80.8</td> |
| </tr> |
| <tr> |
| <td>IFEval</td> |
| <td align="center">80.4</td> |
| <td align="center"><b>82.8</b></td> |
| <td align="center">77.3</td> |
| <td align="center">82.1</td> |
| <td align="center">81.3</td> |
| <td align="center">76.5</td> |
| <td align="center">80.0</td> |
| <td align="center">63.8</td> |
| </tr> |
| <tr> |
| <td align="center" colspan="9" style="background: linear-gradient(90deg, rgba(0,82,204,0.30) 0%, rgba(0,163,255,0.30) 50%, rgba(120,86,255,0.30) 100%); font-weight: bold; height: 32px; padding-top: 2px; padding-bottom: 2px;"><i>Korean Language & Knowledge</i></td> |
| </tr> |
| <tr> |
| <td>KMMLU</td> |
| <td align="center">45.6</td> |
| <td align="center"><b>58.1</b></td> |
| <td align="center">49.9</td> |
| <td align="center">52.0</td> |
| <td align="center">56.4</td> |
| <td align="center">44.4</td> |
| <td align="center">42.7</td> |
| <td align="center">48.5</td> |
| </tr> |
| <tr> |
| <td>CLIcK</td> |
| <td align="center">59.2</td> |
| <td align="center">68.1</td> |
| <td align="center">53.0</td> |
| <td align="center">66.2</td> |
| <td align="center"><b>70.7</b></td> |
| <td align="center">58.2</td> |
| <td align="center">54.7</td> |
| <td align="center">55.0</td> |
| </tr> |
| <tr> |
| <td>HAE-RAE</td> |
| <td align="center">63.0</td> |
| <td align="center">74.0</td> |
| <td align="center">67.0</td> |
| <td align="center">73.7</td> |
| <td align="center"><b>79.2</b></td> |
| <td align="center">70.8</td> |
| <td align="center">55.3</td> |
| <td align="center">56.8</td> |
| </tr> |
| <tr> |
| <td>HRM8K</td> |
| <td align="center"><b>60.4</b></td> |
| <td align="center">46.4</td> |
| <td align="center">51.1</td> |
| <td align="center">49.8</td> |
| <td align="center">51.2</td> |
| <td align="center">31.4</td> |
| <td align="center">30.8</td> |
| <td align="center">54.2</td> |
| </tr> |
| <tr> |
| <td>KoSimpleQA<sup>β </sup></td> |
| <td align="center">71.7</td> |
| <td align="center">77.5</td> |
| <td align="center">45.9</td> |
| <td align="center">78.9</td> |
| <td align="center"><b>88.9</b></td> |
| <td align="center">81.0</td> |
| <td align="center">β</td> |
| <td align="center">72.1</td> |
| </tr> |
| </table> |
| <sub><b>Bold</b> indicates the best score in each row. β indicates not evaluated. <sup>β </sup> KoSimpleQA is evaluated as a judge-free 10-choice MCQA variant. Generative benchmarks use greedy decoding; subtask-style suites report the unweighted subtask mean.</sub> |
| |
| In the broadest same-protocol comparison we are aware of for Korean open models, Rosetta-7B-Instruct holds the **top score on Korean mathematical reasoning (HRM8K 60.4)** β ahead of every Korean flagship in the table β while placing in the leading group on instruction following (IFEval 80.4) and Korean factuality (KoSimpleQA 71.7). Models that lead on knowledge-recall benchmarks are either substantially larger (Midm-2.0-Base, 11.5B) or newest-generation flagships, and none of them match Rosetta on Korean math. |
|
|
| ## Quickstart |
|
|
| ### Transformers |
|
|
| Requires `transformers>=5.13` and `trust_remote_code=True` (the Rosetta architecture ships as custom code in this repository). |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| model_id = "PoSTMEDIA/Rosetta-7B-Instruct" |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, dtype="bfloat16", device_map="auto", trust_remote_code=True |
| ) |
| |
| messages = [ |
| {"role": "user", "content": "νκ΅μ μ ν΅ λ°ν¨ μμ μΈ κ°μ§λ₯Ό μκ°ν΄μ€."}, |
| ] |
| inputs = tokenizer.apply_chat_template( |
| messages, add_generation_prompt=True, return_tensors="pt" |
| ).to(model.device) |
| |
| out = model.generate(inputs, max_new_tokens=1024, temperature=0.7, top_p=0.9, do_sample=True) |
| print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True)) |
| ``` |
|
|
| ### vLLM |
|
|
| Use the PoSTMEDIA vLLM distribution with native Rosetta support β no `trust_remote_code` required: |
|
|
| ```bash |
| VLLM_USE_PRECOMPILED=1 pip install git+https://github.com/PoSTMEDIA-AI/vllm@rosetta-v0.26.0 |
| |
| vllm serve PoSTMEDIA/Rosetta-7B-Instruct --dtype bfloat16 |
| ``` |
|
|
| ```python |
| from openai import OpenAI |
| |
| client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY") |
| resp = client.chat.completions.create( |
| model="PoSTMEDIA/Rosetta-7B-Instruct", |
| messages=[{"role": "user", "content": "λΆμ° μ¬ν 1λ° 2μΌ μ½μ€λ₯Ό μ§μ€."}], |
| temperature=0.7, |
| ) |
| print(resp.choices[0].message.content) |
| ``` |
|
|
| > [!IMPORTANT] |
| > vLLM **v0.26 or later** is required. Recommended sampling: `temperature 0.7, top_p 0.9` (or greedy for deterministic tasks). |
| |
| ## Limitations |
| |
| - The model can generate factually incorrect or outdated information; verify outputs for high-stakes use. |
| - Outputs may reflect biases present in web-scale training data. |
| - Optimized for Korean and English; other languages are not guaranteed. |
| - Alignment was performed on contexts up to 32K tokens; validate quality for longer inputs. |
| |
| ## License |
| |
| Apache License 2.0 β see [LICENSE](./LICENSE). If you build something with Rosetta, we'd appreciate a *"Built with Rosetta"* attribution. |
| |
| ## Citation |
| |
| ```bibtex |
| @misc{rosetta2026, |
| title = {Rosetta-7B: A Bilingual Korean-English Language Model Family}, |
| author = {{PoSTMEDIA AI Lab}}, |
| year = {2026}, |
| url = {https://huggingface.co/collections/PoSTMEDIA/rosetta-6a9db30fd1b4585b0c1845e9} |
| } |
| ``` |
| |
| ## Contact |
| |
| Questions and feedback β please open a discussion on the model page. |
| |
