Llama 3.3:開創合成數據生成新時代

Ai




Llama 3.3 令合成數據生成變得輕而易舉

「我的合成數據成本下降了30倍」

Meta今天推出了Llama 3.3,一款多語言的大型語言模型(LLM),旨在重新定義人工智能在合成數據生成中的角色。Llama 3.3擁有700億個參數,其性能與之前的405B模型相當,但在效率和可及性上進行了優化。

這款多語言模型支持包括印地語、葡萄牙語和泰語等多種語言,使全球開發者能夠為專門的人工智能模型創建定制數據集。

Meta在X平台上表示:「隨著我們繼續探索新的後訓練技術,今天我們推出了Llama 3.3——一款新的開源模型,能在合成數據生成等文本相關用例中提供領先的性能和質量,同時大幅降低推理成本。」

助力合成數據生成

開發者現在可以利用其擴展的上下文長度(128k tokens)來生成龐大且高質量的數據集,解決隱私限制和資源約束等挑戰。

Meta的人工智能主管Yann LeCun曾表示,這一能力使得在低資源語言中的創新成為可能,這一觀點也得到了印度企業家Nandan Nilekani的認同。他提到:「印度應該專注於快速構建小型、特定用例的模型」,強調了Llama在生成印地語模型所需的定制訓練數據方面的重要性。

這種方法的成功在Sarvam AI的Sarvam 2B項目中得到了證明,該項目通過利用Llama生成的合成數據,在印地語任務中超越了更大型的模型。

Meta的機器學習工程師Hamid Shojanazeri表示,合成數據生成能夠解決收集真實世界數據集過於昂貴或不可行的關鍵瓶頸。他補充道:「合成數據對於推進隱私敏感領域或低資源語言的人工智能至關重要。」憑藉其強化學習與人類反饋(RLHF)調整和監督微調,Llama 3.3能夠生成高精度任務所需的指令對齊數據集。

像Sarvam AI和Ola Krutrim等印度初創公司已經開始享受Llama的能力帶來的好處。Sarvam AI的2B模型在2萬億個合成印地語tokens上進行訓練,展示了這類數據如何高效地訓練更小的專用模型,同時保持高性能。

Sarvam AI的首席執行官Vivek Raghavan在接受AIM訪問時表示:「如果你看看印度語言中的1000億tokens,我們使用了一種巧妙的方法來創建合成數據,以便構建這些模型,使用的是Llama 3.1 405B。我們在印度用1,024台NVIDIA H100訓練該模型,僅用了15天。」

同樣,Llama 3.3的多語言支持和可擴展性使其成為彌合未被充分代表語言數據差距的必備工具。

Llama 3.3支持合成數據生成的能力不僅限於小眾用例,還促進了開發者、教育工作者和企業的更廣泛應用。Meta的生成AI副總裁Ahmad Al-Dahle表示:「通過降低生成高質量訓練數據的成本,Llama加速了全球的創新。」

隨著對GPT-4.5的猜測加劇,Llama 3.3明確滿足了開發者的即時需求。憑藉其革命性的合成數據生成和成本效益,Llama 3.3不僅是填補了市場空白,更是在設定新的標準。

KissanAI的聯合創始人Pratik Desai在X上說:「我的合成數據成本下降了30倍。」

為Llama 4奠定基礎

Llama 3.3的發布正好契合Meta的長期人工智能戰略。正如Zuckerberg在Meta的第三季度財報電話會議上透露,預計2025年初推出的Llama 4將引入「新模式、更強的推理能力和更快的性能」。這表明,Llama 3.3中精煉的合成數據生成能力在未來的版本中可能會變得更加強大。

Meta的副總裁Ragavan Srinivasan最近暗示,未來的Llama模型將在「基於記憶的編碼應用和跨模態支持」方面有所進展。Llama 3.3建立的強大框架可能是這些發展的核心。通過使開發者能夠生成特定領域的訓練數據集,Meta將自己定位為公私部門創新的關鍵推動者。

未來的Llama版本可能會支持更廣泛的語言和專門用例。隨著合成數據生成成為人工智能開發的核心,像Llama Guard 3和增強的標記方法將確保安全和負責任的使用。

對於像印度這樣的國家,區域語言的數據創建至關重要,Llama提供了一條可及的途徑來開發與文化相關的人工智能解決方案。

全球範圍內,正如Mark Zuckerberg所提到的,Meta在路易斯安那州的下一代數據中心預示著將推動更雄心勃勃的人工智能進展:「我們是為了長期而來,致力於構建世界上最先進的人工智能。」

這篇文章不僅展示了Llama 3.3在合成數據生成領域的變革性影響,還強調了未來人工智能發展的潛力。隨著技術的進步,開發者和企業將能夠更輕鬆地創建和利用數據,這將為各行各業帶來更多的創新機會。特別是在數據稀缺的地區,這種能力將不僅促進技術的普及,也將推動社會的整體發展。

以上文章由特價GPT API KEY所翻譯及撰寫。而圖片則由FLUX根據內容自動生成。

🎨 Nano Banana Pro 圖像生成器|打幾句說話就出圖

想畫人像、產品圖、插畫?SSFuture 圖像生成器支援 Flux Gemini Nano Banana Pro 改圖 / 合成, 打廣東話都得,仲可以沿用上一張圖繼續微調。

🆓 Flux 模型即玩,不用登入
🤖 登入後解鎖 Gemini 改圖
📷 支援上載參考圖再生成
⚡ 每天免費額度任你玩
✨ 即刻玩 AI 畫圖
Base Setup
keep 100 percent facial information adherence of the attached image and turn her into a lone night hiker standing in the shallow river at the mouth of the canyon, captured as a live action photograph or movie still, not an illustration or CGI render, with an alluring, confident, quietly sexy mood.

Shot and Camera
Three quarter body shot from slightly low height on the riverbank, placing her on the left third while the water leads into the dark canyon and the star filled sky dominates the top of the frame, preserving the original wide vertical composition.

Identity and Pose
She has a slim athletic build, natural proportions, medium length loose hair, and lightly tanned skin, 8k Photorealistic and hyper realistic. She stands barefoot in the water with one leg forward, hips relaxed, wearing a fitted cropped technical top and separate high waisted trail shorts with an open lightweight jacket slipping off one shoulder, one hand loosely holding a headlamp at her side and the other brushing hair from her face as she looks up at the stars.

Lighting and Environment
Keep the soft starlight and faint glow from the distant horizon, giving a gentle rim on her shoulders and hair while a subtle fill from the reflected sky reveals her features. The canyon walls, river reflections, pebbled shore, and dense star field stay exactly like the reference, with small ripples around her feet grounding her in the water.

Masking and Constraints
Change only by adding the subject and her outfit, keep lighting, perspective, white balance, canyon geometry, and river reflections the same, with realistic scale, body proportions, and clear contact between feet and water surface. Change only the requested element and keep lighting, perspective, white balance, pose, face geometry, body proportions, and silhouette the same, and absolutely avoid added text, painterly or toon styling, CGI look, blown highlights, banding, plastic skin, overly tidy backgrounds, or floating feet while requiring consistent perspective, consistent white balance, subtle film like grain, and physically correct contact shadows and reflections. Create a photorealistic and highly detailed image featuring the attached image walking confidently down a modern city street, accompanied by Jason Statham, Dwayne “The Rock” Johnson, and Jason Momoa acting as bodyguards.
John Wick (Keanu Reeves) is walking just beside or slightly behind the subject, holding an umbrella over him to shield from light rain.
The subject should be the central figure, wearing stylish casual clothing — like a fitted jacket, dark jeans, and sunglasses — exuding calm authority and cool charisma.
Statham, The Rock, and Momoa are dressed in black tactical-style suits, maintaining alert, protective stances, scanning the surroundings like professional bodyguards. John Wick wears his signature black suit and tie, looking composed as he holds the umbrella.
The setting is a downtown urban street with wet pavement reflecting city lights, parked luxury cars, and paparazzi in the background snapping photos.
The photo should look like a real paparazzi shot — slightly off-angle, mid-step motion blur, with realistic lighting and reflections.
Lighting: natural daylight with overcast skies, reflections from wet concrete, realistic shadows, subtle raindrops on the umbrella and clothing.
Camera realism: crisp detail on facial features and clothing textures, shallow depth of field emphasizing the group, with lens flare or light bloom for authenticity.
Mood & tone: grounded, cinematic, and stylish — feels like a moment from a celebrity entourage photo or action-movie press capture, taken with an iPhone by paparazzi.
Style: ultra-realistic, documentary-style street photography with modern cinematic sharpness. [Subject]: Young Asian female with "Imada Mio-inspired" doll-like aesthetic (精緻洋娃娃臉). She has large round expressive eyes, a small V-line face, and rosy cheeks. Her expression is innocent, energetic, and slightly flirty. [Hair]: Messy morning hair (剛睡醒的凌亂感), long dark brown hair, slightly tousled, natural volume. [Outfit]: Wearing an oversized translucent white button-down shirt (男友風白襯衫), unbuttoned at the top to reveal collarbones, creating a "bottomless" look (下衣失蹤風格). [Style]: Japanese Gravure Photobook style (寫真集風格), Pure & Sexy vibe, bright high-key lighting, soft skin texture, Fujifilm PRO 400H color tone.

發佈留言

發佈留言必須填寫的電子郵件地址不會公開。 必填欄位標示為 *

🎬 YouTube Premium 家庭 Plan成員一位 只需 HK$148/年

不用提供密碼、不用VPN、無需轉區
直接升級你的香港帳號 ➜ 即享 YouTube + YouTube Music 無廣告播放

立即升級 🔗