Skip to content

Translate Hugging Face blog post: Meta is back with Muse Glimmer: local, agentic, multimodal, and open source - #193

Open
Jwaminju wants to merge 1 commit into
mainfrom
translate/muse-glimmer
Open

Translate Hugging Face blog post: Meta is back with Muse Glimmer: local, agentic, multimodal, and open source#193
Jwaminju wants to merge 1 commit into
mainfrom
translate/muse-glimmer

Conversation

@Jwaminju

Copy link
Copy Markdown
Collaborator

Source: https://huggingface.co/blog/muse-glimmer

This PR adds a Korean translation draft for muse-glimmer.

Downstream handoff:

  • SEO review should use the translation-flow manifest.
  • Quality review should use the translation-flow manifest.

@Jwaminju Jwaminju added the hf-agent:managed Opt PR into HF Agent review automation label Aug 11, 2026
@github-actions

Copy link
Copy Markdown
Contributor
PR Preview Action v1.8.1

🚀 View preview at
https://hugging-face-krew.github.io/pr-preview/pr-193/

Built to branch gh-pages at 2026-08-11 02:27 UTC.
Preview will be ready when the GitHub Pages deployment is complete.

@Jwaminju

Jwaminju commented Aug 11, 2026

Copy link
Copy Markdown
Collaborator Author

HF Agent Review

Gate Result
Quality ❌ Fail
SEO ✅ Pass

Head SHA: 6da4d42b058585a4e579b5c9078013a81e753ec9

Quality report — ❌ Fail

Quality Report

  • Status: reject
  • Quality Score: 0.0
  • Hard failures: 1
  • Issues: 106
  • Source available: True
  • Source changed: False
  • Source segments: 135
  • Target segments: 135

Scorecard

Dimension Score
adequacy 0.0
technical_accuracy 0.0
completeness 0.0
terminology 0.0
fluency 0.0
publishing_integrity 100.0
style_locale 60.0

Metrics

  • qe_metric: heuristic
  • qe_average: 0.9106
  • qe_min: 0.5003
  • embedding_similarity_average: 0.8537
  • embedding_similarity_min: 0.284
  • cache_hits: 8
  • cache_misses: 262

MQM Judge

  • Enabled: True
  • Provider: openai
  • Model: gpt-5.6-luna
  • Reasoning effort: none
  • Prompt: /home/runner/work/hugging-face-krew.github.io/hugging-face-krew.github.io/workflow/skills/quality/judges/mqm_prompt.md
  • Prompt hash: 887d2931aa289213f0bdce4a917a0ac8364dad8e470011069b8f8bf758d69a90
  • Style guide hash: 937d8cd893578d30e716a3eb513cdf5f10d6fd3ad8f5e77068b57f96e160de12
  • Requested segments: 135
  • Evaluated segments: 135
  • MQM errors: 70
  • Cache hits: 0
  • Cache misses: 135
  • Severity counts: {'critical': 1, 'major': 14, 'minor': 55}
  • adequacy_average: 0.9329
  • technical_average: 0.9708
  • fluency_average: 0.9126

Style Guide

  • Enabled: True
  • Guide: /home/runner/work/hugging-face-krew.github.io/hugging-face-krew.github.io/workflow/skills/quality/style/hf-blog-ko-translation-guide.md
  • Policy: /home/runner/work/hugging-face-krew.github.io/hugging-face-krew.github.io/workflow/skills/quality/configs/style_policy.yml
  • Style score: 60.0
  • Rule hits: {'alt_text_caption': 1, 'first_mention_bilingual': 2, 'link_text_translation': 15, 'list_consistency': 1, 'modal_strength': 4, 'translationese': 1}

Style Guide Findings

Rule Severity Segment Current Suggested
translationese minor 에 의해 Rewrite the sentence in natural Korean.
list_consistency minor phrase, phrase, sentence, sentence, sentence Use either sentence-style endings or phrase-style endings consistently within one list.
modal_strength major p_071 또한 다음과 같이 서버에 질의할 수도 있습니다. Preserve the strength of can using: 수 있습니다.
modal_strength major p_072 llama 서버를 Pi와 같은 코딩 에이전트와 함께 사용할 수도 있습니다. Preserve the strength of can using: 수 있습니다.
modal_strength major p_081 다음과 같이 추측 디코딩 드래프터와 함께 llama cli를 사용할 수도 있습니다. Preserve the strength of can using: 수 있습니다.
modal_strength major p_120 참고: Muse Glimmer to Inference Endpoints를 직접 배포하고 에이전트를 연결할 수도 있습니다. Preserve the strength of can using: 수 있습니다.
alt_text_caption minor Deploy Muse Glimmer 30B with Hugging Face Inference Endpoints Translate image alt text while preserving the image path.
link_text_translation minor Check out the demos below for inspiration. Translate link text while preserving the URL target.
link_text_translation minor Muse Glimmer Translate link text while preserving the URL target.
link_text_translation minor as a backbone for various downstream spatial and multimodal tasks Translate link text while preserving the URL target.

Issues

QL-001 technical / major

  • Message: Python/API identifier mismatch.
  • Source: N.B, llama.cpp, llama.cpp, llama.cpp
  • Suggested fix: Preserve source Python/API identifier exactly.
  • Reason: Review gate exact-match validator failed: missing=['N.B', 'llama.cpp', 'llama.cpp', 'llama.cpp']

QL-002 accuracy / major

  • Message: QE metric score is low.
  • Target: 0.5003
  • Suggested fix: Review this segment for omission, unrelated translation, or over-compression.
  • Reason: QE score is below threshold 0.55.

QL-003 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: it’s ideal deploying locally for privacy, reducing costs, or just hacking around.
  • Target: 개인정보 보호를 위해 로컬 배포, 비용 절감, 또는 단순한 해킹에 이상적입니다.
  • Suggested fix: 개인정보를 보호하거나 비용을 절감하거나, 그저 자유롭게 실험해 보기 위해 로컬에서 배포하기에 적합합니다.
  • Reason: ‘로컬 배포, 비용 절감’이 명사구로 병렬되어 문장이 어색하고, ‘just hacking around’의 가벼운 탐색·실험 뉘앙스도 ‘단순한 해킹’으로 부자연스럽게 옮겨졌습니다.

QL-004 terminology / minor

  • Message: MQM judge reported terminology issue.
  • Source: released under the **Apache 2.0 license**
  • Target: **Apache 2.0 license** 하에 공개되어
  • Suggested fix: Apache 2.0 라이선스로 공개되어
  • Reason: 라이선스명은 검색 가능하도록 보존할 수 있지만, 영어 일반명사 ‘license’까지 남겨 한국어 문장이 어색합니다.

QL-005 fluency / major

  • Message: MQM judge reported fluency issue.
  • Source: It’s intended for privacy-aware applications such as coding, document analysis, personal assistants, Claw- or Hermes-like setups.
  • Target: 코딩, 문서 분석, 개인 비서, Claw- 또는 Hermes와 같은 설정과 같은 프라이버시 중심 애플리케이션을 위한 목적입니다.
  • Suggested fix: 코딩, 문서 분석, 개인 비서, Claw나 Hermes와 유사한 설정 등 개인정보 보호를 중시하는 애플리케이션을 대상으로 합니다.
  • Reason: ‘위한 목적입니다’와 ‘같은 설정과 같은’이 중복되고 문법적으로 부자연스러워 독자가 문장 구조를 즉시 파악하기 어렵습니다.

QL-006 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: To celebrate, we are shipping with Meta day-0 support in transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries.
  • Target: 축하의 의미로, transformers, llama.cpp, vLLM, Inference Endpoints 및 기타 라이브러리에서 메타의 day-0 지원을 제공합니다.
  • Suggested fix: 이를 기념해 Transformers, llama.cpp, vLLM, Inference Endpoints 및 기타 라이브러리에서 Meta의 day-0 지원을 제공합니다.
  • Reason: “축하의 의미로”는 영어 구조를 직역한 표현이라 도입부가 어색하고, 공식 라이브러리명인 transformers가 대문자 없이 표기되었습니다. 또한 “메타의 day-0 지원”은 한국어 기술 문맥에서 다소 부자연스럽습니다.

QL-007 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: We built a few cool things and explain our findings in this blog.
  • Target: 우리는 몇 가지 멋진 것들을 만들었고 이 블로그에서 우리의 발견을 설명합니다.
  • Suggested fix: 몇 가지 흥미로운 기능을 구현했으며, 그 결과를 이번 글에서 설명합니다.
  • Reason: “멋진 것들”, “우리의 발견을 설명합니다”는 직역투이며, 기술 블로그 문체에 비해 부자연스럽습니다. 원문의 의미를 유지하면서 “몇 가지 흥미로운 기능을 구현했으며, 그 결과를 이번 글에서 설명합니다”처럼 다듬는 것이 적절합니다.

QL-008 accuracy / major

  • Message: MQM judge reported accuracy issue.
  • Source: **Check out the demos below for inspiration.**
  • Target: **Check out the demos below for inspiration.**
  • Suggested fix: 영감을 얻으려면 아래 데모를 확인해 보세요.
  • Reason: 영어 문장이 그대로 남아 있어 한국어 독자가 데모를 참고하라는 안내를 이해하기 어렵습니다.

QL-009 fluency / major

  • Message: MQM judge reported fluency issue.
  • Source: You can find Muse Glimmer on the Hugging Face Hub.
  • Target: 다음에서 Muse Glimmer on the Hugging Face Hub를 찾을 수 있습니다.
  • Suggested fix: Muse Glimmer는 Hugging Face Hub에서 확인할 수 있습니다.
  • Reason: 영어 전치사 구문인 "on the Hugging Face Hub"가 번역되지 않은 채 문장 중간에 남아 있어 한국어 문장이 부자연스럽고 의미 연결이 깨졌습니다. 또한 원문에는 없는 "다음에서"가 추가되었습니다.

QL-010 accuracy / major

  • Message: MQM judge adequacy score is low.
  • Target: 0.6000
  • Suggested fix: Review this segment against the MQM judge feedback.
  • Reason: MQM adequacy score is below threshold 0.75.

QL-011 style_locale / minor

  • Message: MQM judge reported style_locale issue.
  • Source: <summary>Benchmark results</summary>
  • Target: <summary>Benchmark results</summary>
  • Suggested fix: 벤치마크 결과
  • Reason: HTML 구조는 유지되었지만 summary의 일반 텍스트인 “Benchmark results”가 번역되지 않아 한국어 문서의 언어 일관성이 떨어집니다.

QL-012 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Scores are reported as published. **Bold** indicates the best result among the compared models; ↓ indicates lower is better.
  • Target: 점수는 게시된 방식으로 보고됩니다. **Bold**는 비교 대상 모델 중 최상의 결과를 나타내고; ↓는 더 낮은 값이 더 좋음을 의미합니다.
  • Suggested fix: 점수는 게시된 그대로 표시합니다. 굵게 표시된 값은 비교 대상 모델 중 가장 좋은 결과를 나타내며, ↓는 값이 낮을수록 좋다는 의미입니다.
  • Reason: ‘게시된 방식으로 보고됩니다’는 published의 의미를 어색하고 불명확하게 옮겼으며, 세미콜론 앞뒤의 문장 연결도 한국어에서 부자연스럽습니다. 또한 표의 서식 표기인 ‘Bold’를 그대로 두기보다 굵게 표시된 항목을 설명하는 표현이 명확합니다.

QL-013 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Agentic Coding
  • Target: 에이전트적 코딩
  • Suggested fix: Agentic Coding
  • Reason: 의미는 이해되지만 ‘에이전트적 코딩’은 한국어 기술 문맥에서 다소 부자연스럽습니다. ‘Agentic Coding’은 검색성과 용어 인지도를 위해 원문을 유지하거나 자연스러운 설명형 표현을 사용하는 편이 적절합니다.

QL-014 terminology / minor

  • Message: MQM judge reported terminology issue.
  • Source: Agentic Coding
  • Target: 에이전트적 코딩
  • Suggested fix: 에이전트 코딩(Agentic Coding)
  • Reason: Agentic Coding을 직역한 ‘에이전트적 코딩’은 한국어 기술 문맥에서 부자연스럽고 검색성이 낮습니다. Agentic coding의 의미를 유지하면서 통용되는 표현이나 원문 병기를 사용하는 편이 적절합니다.

QL-015 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Agentic Coding
  • Target: 에이전트적 코딩
  • Suggested fix: 에이전트 코딩
  • Reason: 기술 블로그 표의 항목명으로 ‘에이전트적 코딩’은 한국어에서 다소 직역투이고 부자연스럽습니다. 원문의 agentic coding 개념을 검색성과 자연스러움을 고려해 ‘에이전트 코딩’ 또는 ‘에이전트 기반 코딩’으로 옮기는 편이 적절합니다.

QL-016 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Agentic Coding
  • Target: 에이전트적 코딩
  • Suggested fix: 에이전트 코딩
  • Reason: 기술 표의 항목명으로 ‘에이전트적 코딩’은 의미는 이해되지만 한국어에서 부자연스럽고 통용성이 낮습니다. 원문의 기술 개념을 보존하면서 검색성과 자연스러움을 높이려면 ‘에이전트 코딩’ 또는 ‘에이전트 기반 코딩’이 적절합니다.

QL-017 terminology / minor

  • Message: MQM judge reported terminology issue.
  • Source: Attack Success Rate (↓): 28.4<br>Utility: 94.2
  • Target: Attack Success Rate (↓): 28.4<br>Utility: 94.2
  • Suggested fix: 공격 성공률(↓): 28.4
    효용성: 94.2
  • Reason: 표의 일반 기술 용어인 Attack Success Rate와 Utility가 번역되지 않아 한국어 표 안에서 검색성과 가독성이 낮아집니다. 수치, 하강 방향 표시, 줄바꿈은 보존되어야 합니다.

QL-018 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Muse Glimmer is a dense 30B parameter model consisting of:
  • Target: Muse Glimmer는 30B 파라미터의 밀도형 모델로 구성되어 있습니다:
  • Suggested fix: Muse Glimmer는 30B 파라미터의 밀집형 모델이며, 다음으로 구성됩니다:
  • Reason: dense model을 ‘밀도형 모델’로 옮긴 표현은 한국어 기술 문맥에서 부자연스럽고, ‘구성되어 있습니다’도 원문의 모델 설명과 어울리지 않습니다. 다음 항목을 소개하는 문장이므로 자연스럽게 ‘밀집형 30B 파라미터 모델이며, 다음으로 구성됩니다’처럼 표현하는 편이 적절합니다.

QL-019 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: in exchange for some memory cost
  • Target: 약간의 메모리 비용과 교환으로
  • Suggested fix: 메모리를 일부 더 사용하는 대신
  • Reason: 의미는 전달되지만 ‘메모리 비용과 교환으로’는 영어식 표현이라 한국어 기술 문장으로 다소 어색합니다.

QL-020 accuracy / major

  • Message: MQM judge reported accuracy issue.
  • Source: Each key-value head is shared by 16 query heads
  • Target: 각 키-값 헤드는 16개의 쿼리 헤드가 공유하여
  • Suggested fix: 각 키-값 헤드는 16개의 쿼리 헤드가 공유하며
  • Reason: 원문은 각 key-value head 하나를 16개의 query head가 공유한다는 구조인데, 번역문은 주어와 조사가 어색하게 결합되어 16개의 query head가 key-value head를 공유한다는 관계가 명확하지 않고 문법적으로도 부자연스럽습니다.

QL-021 accuracy / major

  • Message: MQM judge reported accuracy issue.
  • Source: reduces KV-cache memory by 16x
  • Target: KV-캐시 메모리를 16배로 줄이고
  • Suggested fix: KV-캐시 메모리를 16분의 1로 줄이고
  • Reason: 원문의 16x 감소는 메모리를 16분의 1로 줄인다는 의미인데, '16배로 줄이고'는 한국어에서 16배 감소 또는 16배만큼 감소로 해석될 수 있어 수치 의미가 모호하거나 잘못 전달될 수 있습니다.

QL-022 technical / major

  • Message: MQM judge technical score is low.
  • Target: 0.7200
  • Suggested fix: Review this segment against the MQM judge feedback.
  • Reason: MQM technical score is below threshold 0.75.

QL-023 accuracy / major

  • Message: MQM judge reported accuracy issue.
  • Source: Perception Encoder was previously introduced by Meta as a backbone for various downstream spatial and multimodal tasks.
  • Target: Perception Encoder는 이전에 Meta as a backbone for various downstream spatial and multimodal tasks에 의해 소개되었습니다.
  • Suggested fix: Perception Encoder는 다양한 다운스트림 공간 및 멀티모달 작업을 위한 백본으로 Meta가 이전에 소개한 모델입니다.
  • Reason: 문장의 핵심 서술부인 '다양한 다운스트림 공간 및 멀티모달 작업을 위한 백본으로 Meta가 이전에 소개했다'가 영어 원문 그대로 삽입되어 한국어 문법과 의미가 완성되지 않았습니다. 또한 'as ...에 의해' 구조로 인해 주체와 용도가 부자연스럽고 명확하게 전달되지 않습니다.

QL-024 accuracy / major

  • Message: MQM judge reported accuracy issue.
  • Source: The encoder patchifies images to a shape of 2 frames x 3 channels x 14 x 14, and passes them through a linear layer for projection. An interpolated absolute position embedding from a learned position table is then added to these embeddings.
  • Target: 인코더는 이미지를 2 프레임 x 3 채널 x 14 x 14의 형태로 패치화하고, 이를 학습된 위치 테이블에서 보간된 절대 위치 임베딩을 더한 뒤, 이 임베딩들을 시각 타워로 전달합니다.
  • Suggested fix: 인코더는 이미지를 2 frames x 3 channels x 14 x 14 형태로 패치화한 뒤, 프로젝션을 위해 선형 계층에 통과시킵니다. 그런 다음 학습된 위치 테이블에서 보간한 절대 위치 임베딩을 이러한 임베딩에 더하고, 결과를 비전 타워로 전달합니다.
  • Reason: 원문은 패치화한 이미지 임베딩을 선형 계층에 통과시켜 프로젝션한 다음, 학습된 위치 테이블에서 보간한 절대 위치 임베딩을 더한다고 설명합니다. 대상 문장에서는 선형 계층을 통한 프로젝션 단계가 누락되었고, 위치 임베딩을 더하는 대상과 시각 타워로 보내는 순서도 합쳐져 기술적 처리 흐름이 달라졌습니다.

QL-025 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: These are then sent to the vision tower which consist of 50 layers and GELU MLPs.
  • Target: 시각 타워는 50개의 계층과 GELU MLP로 구성됩니다.
  • Suggested fix: 이 임베딩들은 50개 레이어와 GELU MLP로 구성된 비전 타워로 전달됩니다.
  • Reason: 의미는 대체로 전달되지만 vision tower를 앞 문장의 '시각 타워'와 연결하면서 원문의 '이 임베딩들이 전달된다'는 지시 대상이 사라졌습니다. 또한 기술 문맥에서는 '비전 타워'가 더 일관되고 자연스럽습니다.

QL-026 technical / major

  • Message: MQM judge technical score is low.
  • Target: 0.7200
  • Suggested fix: Review this segment against the MQM judge feedback.
  • Reason: MQM technical score is below threshold 0.75.

QL-027 accuracy / major

  • Message: MQM judge reported accuracy issue.
  • Source: caps the clip at 96 frames sampled evenly across video
  • Target: 비디오를 고르게 샘플링하여 96프레임으로 상한을 둡니다
  • Suggested fix: 비디오 전체에서 프레임을 고르게 샘플링하며 클립은 최대 96프레임으로 제한합니다
  • Reason: 원문은 비디오 전체에 걸쳐 균등하게 샘플링한 프레임 수가 최대 96개라는 뜻입니다. 현재 번역은 ‘비디오를 96프레임으로 샘플링한다’는 의미로 읽혀, 클립(clip)에 대한 상한과 샘플링 대상의 관계가 불명확하고 기술적 조건을 정확히 전달하지 못합니다.

QL-028 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: interleaving text with frame e.g. “Time: 0.0s <|video|> x N”
  • Target: 예를 들어 텍스트와 프레임을 교차시키며 “Time: 0.0s <|video|> x N” 형태를 삽입합니다
  • Suggested fix: 예를 들어 “Time: 0.0s <|video|> x N”과 같은 형식으로 텍스트 사이에 프레임 플레이스홀더를 교차 배치합니다
  • Reason: 원문의 ‘텍스트와 프레임을 교차 배치한다’는 구조가 현재 번역에서는 ‘형태를 삽입한다’로 이어져 무엇을 삽입하는지 다소 어색하고 모호합니다.

QL-029 technical / critical

  • Message: MQM judge reported technical issue.
  • Source: Transformers
  • Target: 트랜스포머
  • Suggested fix: Transformers
  • Reason: Hugging Face의 라이브러리명인 Transformers를 일반 명사처럼 음역해 고유 제품명이 사라졌습니다. 기술 문서에서 검색성과 정확한 라이브러리 식별을 해치며, 라이브러리명은 원문 표기를 유지해야 합니다.

QL-030 accuracy / major

  • Message: MQM judge adequacy score is low.
  • Target: 0.4000
  • Suggested fix: Review this segment against the MQM judge feedback.
  • Reason: MQM adequacy score is below threshold 0.75.

QL-031 terminology / major

  • Message: MQM judge reported terminology issue.
  • Source: transformers
  • Target: 트랜스포머
  • Suggested fix: Transformers
  • Reason: Hugging Face 라이브러리명인 transformers를 일반 명사처럼 번역해 제품·라이브러리 식별자가 불명확해졌습니다.

QL-032 terminology / minor

  • Message: MQM judge reported terminology issue.
  • Source: in transformers
  • Target: 트랜스포머에서
  • Suggested fix: Transformers에서
  • Reason: 라이브러리명 transformers가 일반 명사처럼 한글화되어 검색성이 떨어집니다. 또한 원문의 ‘in Transformers’는 라이브러리에서의 지원을 뜻하므로 제품명을 보존하는 것이 적절합니다.

QL-033 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: both for the main model and the speculative decoding drafter
  • Target: 주요 모델과 추측 디코더 드래프터 모두에 대해 그렇습니다.
  • Suggested fix: 주요 모델과 추측 디코딩 드래프터 모두를 지원합니다.
  • Reason: 앞 문장을 불필요하게 분리한 뒤 ‘그렇습니다’로 지시해 문장이 어색하고 무엇을 수식하는지 즉시 파악하기 어렵습니다. 원문의 범위를 한 문장 안에서 자연스럽게 연결하는 편이 명확합니다.

QL-034 accuracy / minor

  • Message: MQM judge reported accuracy issue.
  • Source: device_map="auto" places the model on whichever accelerator is available.
  • Target: device_map="auto"은 사용 가능한 가속기에 모델을 배치합니다.
  • Suggested fix: device_map="auto"는 사용 가능한 가속기 중 하나에 모델을 배치합니다.
  • Reason: 원문의 ‘whichever accelerator is available’은 사용 가능한 가속기 중 어느 것이든 자동으로 선택한다는 의미인데, 번역문은 ‘사용 가능한 가속기에’로 옮겨 자동 선택의 의미가 다소 약해졌습니다. 또한 인라인 코드 뒤 조사는 기술 문서에서 자연스럽게 붙이려면 ‘는’이 적절합니다.

QL-035 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: torchvision
  • Target: torchvision가
  • Suggested fix: 이미지와 텍스트를 사용하려면 torchvision이 필요합니다.
  • Reason: 라이브러리명 뒤에 ‘가’를 붙일 때는 받침이 있는 외래어 표기 관습상 ‘torchvision이’가 자연스럽습니다. 의미와 기술 정보에는 영향이 없습니다.

QL-036 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Muse Glimmer accepts images as input, as demonstrated here:
  • Target: Muse Glimmer는 입력으로 이미지를 허용합니다. 아래에서 시연합니다:
  • Suggested fix: Muse Glimmer는 아래 예시처럼 이미지를 입력으로 받을 수 있습니다:
  • Reason: 의미는 전달되지만 ‘입력으로 이미지를 허용합니다’와 ‘아래에서 시연합니다’는 영어식이고 다소 어색합니다. 원문의 ‘as demonstrated here’는 아래 예시에서 확인할 수 있다는 뜻으로 자연스럽게 연결하는 편이 좋습니다.

QL-037 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: You can do video inference as follows, here’s an example from VideoMME2, which is the most popular video question answering benchmark.
  • Target: 아래와 같이 비디오 추론을 수행할 수 있으며, 아래 예시는 VideoMME2로, 가장 인기 있는 비디오 질문 응답 벤치마크입니다.
  • Suggested fix: 다음과 같이 비디오 추론을 수행할 수 있습니다. 아래는 가장 널리 사용되는 비디오 질의응답 벤치마크인 VideoMME2의 예시입니다.
  • Reason: ‘아래’가 반복되고 ‘예시는 VideoMME2로’가 자연스럽지 않아 문장 흐름이 어색합니다. 원문의 예시 소개와 벤치마크 설명은 유지하면서 한국어 문장으로 자연스럽게 연결해야 합니다.

QL-038 accuracy / minor

  • Message: MQM judge reported accuracy issue.
  • Source: here’s how you can do it.
  • Target: Muse Glimmer는 멀티모달 도구 호출을 수행할 수 있습니다. 아래 예시에서는 이미지 속 도시를 기반으로 날씨 도구를 호출하도록 모델에 요청합니다.
  • Suggested fix: Muse Glimmer는 멀티모달 도구 호출을 수행할 수 있습니다. 이제 이를 수행하는 방법을 살펴보겠습니다. 아래 예시에서는 이미지 속 도시를 기반으로 날씨 도구를 호출하도록 모델에 요청합니다.
  • Reason: 원문의 ‘여기서는 이를 수행하는 방법을 보여줍니다’라는 안내 문장이 번역에서 누락되었습니다. 뒤 문장이 예시를 소개하기는 하지만, 원문의 방법 안내 의미가 직접적으로 전달되지 않습니다.

QL-039 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: open ended object detection
  • Target: 열린(Open-ended) 객체 검출
  • Suggested fix: 개방형 객체 검출
  • Reason: ‘open ended’를 ‘열린’으로 직역하면 한국어 기술 문맥에서 부자연스럽고 의미가 불분명합니다. ‘개방형 객체 검출’ 또는 ‘오픈셋 객체 검출’처럼 문맥에 맞는 표현이 더 명확합니다. 괄호 안의 영문도 원문에 없는 불필요한 병기입니다.

QL-040 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Here is an end to end script to perform object detection GitHub Gist
  • Target: 여기에 객체 검출을 수행하는 엔드 투 엔드 스크립트가 있습니다 GitHub Gist
  • Suggested fix: 객체 검출을 수행하는 엔드투엔드(end-to-end) 스크립트는 GitHub Gist에서 확인할 수 있습니다.
  • Reason: 문장 끝의 GitHub Gist가 앞부분과 문법적으로 연결되지 않아 문장이 어색하고, 원문의 링크 또는 출처 표기 구조가 불분명합니다. 또한 end to end는 가이드의 권장 표기인 엔드투엔드(end-to-end)가 아닙니다.

QL-041 accuracy / major

  • Message: MQM judge adequacy score is low.
  • Target: 0.7200
  • Suggested fix: Review this segment against the MQM judge feedback.
  • Reason: MQM adequacy score is below threshold 0.75.

QL-042 terminology / minor

  • Message: MQM judge reported terminology issue.
  • Source: calibrated quants
  • Target: 보정된 양자값
  • Suggested fix: 보정된 양자화 버전(calibrated quants)
  • Reason: 기술 용어 quants를 ‘양자값’으로만 옮겨 검색성과 기술 문맥이 떨어집니다. 이 문맥에서는 양자화 파일 또는 양자화 버전으로 표현하고 원어를 병기하는 것이 명확합니다.

QL-043 technical / minor

  • Message: MQM judge reported technical issue.
  • Source: this repo
  • Target: this repo
  • Suggested fix: 이 저장소
  • Reason: 문장 안에 generic English 표현이 그대로 남아 있어 한국어 문체와 용어 처리가 일관되지 않습니다. 기술적 의미를 보존하면서 ‘이 저장소’로 번역하는 것이 자연스럽습니다.

QL-044 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: run:
  • Target: 실행하세요:
  • Suggested fix: 다음 명령을 실행합니다:
  • Reason: 공식 기술 블로그 문체에서 ‘실행하세요’는 해요체로 다소 가볍고, 콜론 뒤 명령 안내와도 어색하게 연결됩니다.

QL-045 accuracy / major

  • Message: MQM judge reported accuracy issue.
  • Source: to provide same output with extra speed-ups in decoding phase
  • Target: 디코딩 단계에서 추가 속도 향상을 제공하기 위해
  • Suggested fix: 동일한 출력을 유지하면서 디코딩 단계에서 추가 속도 향상을 제공하기 위해
  • Reason: 원문의 핵심 주장인 '동일한 출력(same output)'이 번역에서 누락되었습니다. 따라서 DFlash가 같은 출력을 유지하면서 디코딩 속도를 추가로 높인다는 의미가 전달되지 않습니다.

QL-046 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: infer like how you would with base model with an additional parameter (shown in the upcoming snippets)
  • Target: 기본 모델에서처럼 추가 매개변수를 사용하여 추론할 수 있습니다(다음 코드 조각에서 보여줍니다)
  • Suggested fix: 기본 모델을 사용할 때와 같은 방식으로 추가 매개변수를 전달해 추론할 수 있습니다(아래 코드 조각에서 확인할 수 있습니다).
  • Reason: 의미는 전달되지만 ‘기본 모델에서처럼’은 한국어로 다소 어색하고, 괄호 안의 ‘보여줍니다’는 무엇을 보여주는지 불분명합니다. 원문의 ‘기본 모델을 사용할 때와 같은 방식으로’라는 흐름을 자연스럽게 풀어 쓰는 편이 좋습니다.

QL-047 terminology / minor

  • Message: MQM judge reported terminology issue.
  • Source: llama cli
  • Target: llama cli
  • Suggested fix: 공식 명칭이 llama-cli인 경우 다음과 같이 추측 디코딩 드래프터와 함께 llama-cli를 사용할 수도 있습니다.로 표기
  • Reason: 의미는 전달되지만 CLI 명칭은 일반적으로 llama-cli처럼 공식 표기를 확인해 유지하는 것이 검색성과 기술 정확성에 더 적합합니다. 다만 제공된 원문 표기 자체가 llama cli이므로 심각한 오류는 아닙니다.

QL-048 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Muse Glimmer 30B Inference Endpoints preset
  • Target: Muse Glimmer 30B Inference Endpoints preset
  • Suggested fix: Muse Glimmer 30B Inference Endpoints 프리셋
  • Reason: 제품명과 기능명은 보존했지만 영어 일반 명사인 preset까지 그대로 두어 한국어 문장 안에서 다소 어색합니다. 검색 가능한 고유명은 유지하면서 일반 명사는 한국어로 옮기는 편이 자연스럽습니다.

QL-049 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: The deployed model exposes an OpenAI-compatible Chat Completions API.
  • Target: 배포된 모델은 OpenAI-호환 Chat Completions API를 노출합니다.
  • Suggested fix: 배포된 모델은 OpenAI 호환 Chat Completions API를 제공합니다.
  • Reason: 의미와 기술 명칭은 보존되었지만, ‘API를 노출합니다’는 영어식 표현으로 한국어 기술 문서에서 다소 어색합니다.

QL-050 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: See the Inference Endpoints documentation for configuration, autoscaling, security, logs, and monitoring.
  • Target: 구성, 자동 확장, 보안, 로그 및 모니터링에 대한 내용은 Inference Endpoints documentation를 참고하십시오.
  • Suggested fix: 구성, 자동 확장, 보안, 로그 및 모니터링에 대한 내용은 Inference Endpoints 문서를 참고하세요.
  • Reason: 제품명인 Inference Endpoints는 유지할 수 있지만 일반 명사 documentation까지 영어로 남아 있고 한국어 조사와 직접 결합되어 문장이 부자연스럽습니다.

QL-051 technical / minor

  • Message: MQM judge reported technical issue.
  • Source: on bf16 with Hopper-class GPUs with 80GB VRAM each
  • Target: Hopper급 GPU 80GB VRAM 각각으로 bf16에서
  • Suggested fix: 각각 80GB VRAM을 탑재한 Hopper급 GPU에서 bf16으로
  • Reason: 원문의 ‘각 GPU가 80GB VRAM을 보유한다’는 조건이 현재 표현에서는 어색하고 수식 관계가 불명확합니다. 의미는 추론할 수 있지만 실험 조건을 정확히 전달하려면 GPU마다 80GB VRAM이 있었다는 구조를 명시하는 것이 좋습니다.

QL-052 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: Workload | Practical minimum
  • Target: 작업 부하 | 실용적 최소값
  • Suggested fix: 워크로드 | 현실적인 최소 사양
  • Reason: 표 구조와 의미는 대체로 보존되었지만, 기술 문맥에서 'workload'는 일반적으로 '워크로드'로 옮기며 'practical minimum'을 '실용적 최소값'으로 직역하면 한국어 표 머리글로 다소 어색합니다.

QL-053 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: slow/tight
  • Target: 느림/타이트
  • Suggested fix: 느리고 여유가 적음
  • Reason: 의미는 대체로 전달되지만, 표의 기술적 제약을 설명하는 표현으로 ‘느림/타이트’는 한국어 문맥에서 부자연스럽고 모호합니다. ‘느리고 여유가 적음’처럼 성능과 메모리 여유 부족을 명확히 표현하는 것이 읽기 쉽습니다.

QL-054 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: 8 GPUs is usually insufficient
  • Target: 8 GPUs는 보통 부족
  • Suggested fix: 8 GPUs로는 보통 부족합니다
  • Reason: 의미는 전달되지만 서술어가 생략된 명사구 형태가 표의 설명으로는 다소 어색하고 문장이 완결되지 않았습니다.

QL-055 accuracy / major

  • Message: MQM judge reported accuracy issue.
  • Source: fine-tune Muse Glimmer on small split of MolmoWeb dataset
  • Target: Muse Glimmer on small split of MolmoWeb dataset를 파인튜닝하는
  • Suggested fix: MolmoWeb 데이터셋의 작은 분할을 사용해 Muse Glimmer를 미세 조정하는
  • Reason: 원문의 fine-tune 대상인 Muse Glimmer와 데이터셋의 small split 조건이 한국어 문장 안에서 분리되지 않고 영어 구문 그대로 남아 의미 관계와 데이터셋 조건을 명확히 전달하지 못합니다.

QL-056 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: This shows how to make model generate structured outputs and how to fine-tune on images.
  • Target: 이는 모델이 구조화된 출력을 생성하도록 만드는 방법과 이미지를 대상으로 파인튜닝하는 방법을 보여줍니다.
  • Suggested fix: 이 예제에서는 모델이 구조화된 출력을 생성하도록 하는 방법과 이미지로 미세 조정(fine-tuning)하는 방법을 보여줍니다.
  • Reason: make model generate의 목적어와 주어 관계가 다소 직역투이며, fine-tune을 파인튜닝으로 옮겨 용어 가이드의 첫 등장 권장 표현인 미세 조정(fine-tuning)과 어긋납니다.

QL-057 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: running the model on OpenCode with AsyncGRPO example
  • Target: OpenCode with AsyncGRPO example에서 모델을 실행하는 실험도 수행했습니다
  • Suggested fix: OpenCode의 AsyncGRPO 예제에서 모델을 실행해 보는 실험도 진행했습니다.
  • Reason: 예제 이름을 영어 문장 구조 그대로 붙여 한국어 문장이 어색하고, 무엇이 예제의 명칭인지 명확하지 않습니다. 기술적 의미는 대체로 전달되지만 블로그 문체에 맞게 예제명을 명확히 처리하는 것이 좋습니다.

QL-058 fluency / minor

  • Message: MQM judge reported fluency issue.
  • Source: try training with coding environments in OpenEnv and TRL
  • Target: OpenEnv 및 TRL에서 코딩 환경으로의 훈련을 시도해 보시길 권장합니다
  • Suggested fix: OpenEnv와 TRL의 코딩 환경을 활용해 학습을 시도해 보시길 권장합니다.
  • Reason: ‘코딩 환경으로의 훈련’은 원문의 training with coding environments를 직역한 표현으로, 한국어 기술 문맥에서 의미가 부자연스럽습니다. OpenEnv와 TRL을 사용해 코딩 환경에서 학습을 시도한다는 뜻으로 풀어 쓰는 편이 명확합니
SEO report — ✅ Pass

SEO Eval Report

Gate: ✅ PASS — deterministic AND rubric

  • File: ../target/_posts/2026-08-10-muse-glimmer.md
  • Source: —
  • Primary keyword: (none — D5 skipped)
  • Mode: file

Gate

  • Status: PASS
  • Blockers: ✅ pass
  • Deterministic REQUIRED (D1–D7): ✅ pass
  • Rubric (R1–R6): ✅ pass (mean None, min None)

Blockers

✅ body_not_empty: Body is not empty
✅ robots_indexable: Robots is indexable
✅ internal_links_resolve: All internal links resolve
✅ local_images_resolve: All local images resolve

Required checks (gated)

✅ heading_hierarchy: Heading hierarchy: Valid
✅ alt_text_coverage: Alt text coverage: 1/1 images
✅ descriptive_alt_text: Descriptive alt text: 1/1 (≥80% recommend)
✅ image_files_exist: All 0 local image file(s) exist

OpenAI rubric checks

✅ semantic_metadata: PASS (required) — Missing meta description in frontmatter (description is empty).
✅ alt_semantics: PASS (review) — https://endpoints.huggingface.co/social-share/thumbnail-main.jpg: Non-empty; describes deploying Muse Glimmer 30B with Hugging Face Inference Endpoints.

Advisory checks (not gated)

✅ opening_summary: Opening 3 paragraphs: 460 chars (recommend ≥150 for KO/GEO)
✅ h1_count: Markdown H1 count: 1 (review against rendered layout)
✅ citations: Citations/statistics: 25 (recommend ≥1 for GEO)
⚠️ question_headings: Scannable H2/H3 (question or keyword): 0 (0 question, 0 keyword)
⚠️ internal_links: Internal links: 1 (recommend 2-3)
✅ word_count: Word count: 3303 (recommend ≥300)
ℹ️ primary_keyword: No primary_keyword in manifest — keyword check skipped
⚠️ webp_format: WebP format: 0/1 images (≥50% recommend)
⚠️ lazy_loading: Lazy loading: 0 images (optional)

Signals (evidence — not directly gated)

  • Frontmatter: title 53 chars, description 0 chars, author present True
  • Title text: Meta가 Muse Glimmer로 돌아왔다: 로컬, 에이전트형, 멀티모달, 그리고 오픈 소스!
  • Description text: —
  • Opening text: * TOC
    {:toc}

이 글은 Hugging Face 블로그의 Meta is back with Muse Glimmer: local, agentic, multimodal, and open source를 한국어로 번역한 글입니다.

  • Opening: first paragraph 184 chars, first 3 paragraphs 460 chars
  • Headings: markdown H1 1, rendered effective H1 2, layout title H1 True
  • Links: total 24, external 23, internal 0, citation signals 25
  • Images: total 1, empty alt 0, filename-like alt 0, missing local files 0

Semantic review packet

  • Title: Meta가 Muse Glimmer로 돌아왔다: 로컬, 에이전트형, 멀티모달, 그리고 오픈 소스!
  • Description: —
  • Rendered H1 candidates: Meta가 Muse Glimmer로 돌아왔다: 로컬, 에이전트형, 멀티모달, 그리고 오픈 소스!, Meta가 Muse Glimmer로 돌아왔다: 로컬, 에이전트형, 멀티모달, 그리고 오픈 소스!
  • Opening: * TOC
    {:toc}

이 글은 Hugging Face 블로그의 Meta is back with Muse Glimmer: local, agentic, multimodal, and open source를 한국어로 번역한 글입니다.

  • Canonical/permalink: —
  • Instruction: Compare title, description, rendered H1, and opening text for meaning consistency. This packet is evidence only; it does not decide pass/fail.

Frontmatter (advisory — written by metadata step, not gated)

✅ title: Title: 53 chars (recommend ≤60)
❌ description: Description is missing
✅ image: OG image: assets/images/blog/posts/2026-08-10-muse-glimmer/thumbnail.png
✅ categories: Categories: 2 (recommend 2-3)
✅ author: Author: dailybot

SEO metadata suggestion — PARTIAL

This is a suggestion. SEO is applied only when the post frontmatter is updated.
To apply safe fields from a partial suggestion, leave a trusted PR comment: metadata apply.

  • Auto apply: False
  • Requires human: True
  • Mode: frontmatter_only
  • Reason: metadata candidate needs policy decisions or missing title/description

Candidate

  • title: Meta가 Muse Glimmer로 돌아왔다: 로컬, 에이전트형, 멀티모달, 그리고 오픈 소스!
  • categories: ['Translation', 'HuggingFace']
  • image: assets/images/blog/posts/2026-08-10-muse-glimmer/thumbnail.png

Needs policy decision

  • target_url
  • source_url
  • canonical_policy
  • translation_indexing
  • target_locale
  • source_locale

Warnings

  • None

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

hf-agent:managed Opt PR into HF Agent review automation

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant