ChatML (Chat Markup Language) la dinh dang van ban dung de bieu dien cuoc tro chuyen giua nguoi dung va AI. Noi don gian hon, day la cach chung ta viet lai mot cuoc chat thanh cac doan van ban co nhan (role) ro rang, de model AI hieu duoc ai dang noi va noi gi.
ChatML La Gi?
ChatML la mot chuan dinh dang van ban (text-based format) de truyen thong tin giua nguoi dung va AI thanh cac luot luot (turns) co cau truc. Moi luot duoc gan mot vai tro (role) nhat dinh, thuong la system, user, hoac assistant.
Khac voi cac giao thuc phuc tap nhu JSON hay XML, ChatML chi la van binh thuong co them cac token danh dau. Vi du, ban co the viet mot tin nhan nhu:
<|im_start|>system
Ban la mot tro ly AI huu ich.
<|im_end|>
<|im_start|>user
ChatML la gi?
<|im_end|>
<|im_start|>assistant
ChatML la dinh dang...
<|im_end|>Nhu ban thay, moi doan bat dau bang <|im_start|> va ket thuc bang <|im_end|>. Day la cach model phan biet noi dung cua tung ben trong cuoc tro chuyen.
Tai Sao ChatML Ra Doi?
Truoc ChatML, cac model AI nhan dau vao nhu mot doan van ban lien tuc. Nghia la neu ban muon hoi model mot cau hoi, ban chi don gian noi dung dau vao vao. Khong co cach nao de model biet phan nao la huong dan, phan nao la cau hoi cua nguoi dung.
Dieu nay gay ra nhieu van de. Model khong phan biet duoc noi dung nao la lenh cua nguoi lap trinh, noi dung nao la cau hoi cua nguoi dung cuoi. Ket qua la model de bi nhan nhieu lenh khong mong muon, hoac tra loi khong dung ngu canh.
ChatML giai quyet bang cach dua vao cac special tokens (token dac biet) de tach biet tung phan. Model duoc huan luyen de nhan biet cac token nay va xu ly cuoc tro chuyen theo dung ngu canh.
Cau Truc Cua ChatML Nhu The Nao?
Mot doan ChatML thuong gom ba loai vai tro chinh:
1. System (He thong): Day la phan huong dan tong quat, noi cho model biet no dang dong vai tro gi. Vi du: “Ban la mot chuyen gia SEO” hoac “Tra loi bang tieng Viet, go ngan gon”. Phan system thuong xuat hien mot lan o dau cuoc tro chuyen.
2. User (Nguoi dung): Day la cau hoi hoac yeu cau cua nguoi dung. Moi lan nguoi dung hoi mot cau, no duoc boc lai trong mot khoi <|im_start|>user den <|im_end|>.
3. Assistant (Tro ly AI): Day la phan tra loi cua model. Khi model tao ra van ban, no cung xuat ra cac token danh dau tuong tu de phan biet phan tra loi cua minh.
Mot so dinh dang ChatML mo rong con ho tro tool hoac function role, cho phep model goi cac cong cu ben ngoai trong luc tra loi.
ChatML Khac Gi Voi Cac Dinh Dang Khac?
Co nhieu cach de bieu dien cuoc tro chuyen cho AI. OpenAI dung Chat Completions API voi JSON (gom cac truong role va content). Anthropic dung mot dinh dang tuong tu nhung them truong system rieng biet. Mistral va Google Gemini cung co cach cau truc rieng.
ChatML noi bat o cho no la dinh dang van ban thuan, khong phai JSON. Dieu nay co y nghia lon khi ban chay model local. Vi ChatML chi la van ban, ban khong can mot parser JSON phuc tap de xu ly. Model chi can doc dau vao nhu mot chuoi ky tu thong thuong.
Dieu nay cung giup ChatML hieu qua hon ve mat token. Khong co cac dau ngoac kep, dau phay, hay ten truong lap diem lap lai nhu trong JSON. Moi doan chat chi ton them may token cho <|im_start|> va <|im_end|>.
Ai Dang Su Dung ChatML?
ChatML duoc cong bo lan dau boi OpenAI nam 2022, dung cho model nhu GPT-3.5 va GPT-4. Sau do, nhieu nha phat trien model lay yeu tuong nay de ap dung cho model cua minh.
Qwen cua Alibaba la mot trong nhung model open-weight noi nhat su dung dinh dang ChatML. Cac model nhu Qwen2, Qwen2.5, Qwen3 deu su dung ChatML lam dinh dang giao tiep chinh. Noi cach khac, neu ban chay Qwen local bang llama.cpp, Ollama, hay vLLM, ban dang dung ChatML ma khong can biet.
Mot so model khac nhu Yi, DeepSeek, hay CodeLlama cung su dung cac bien the gan giong ChatML, chi khac nhau o cu phap token danh dau.
Cach Su Dung ChatML Trong Thuc Te
Neu ban dung Ollama hay llama.cpp de chay model local, ban da tuong tac voi ChatML thong qua cac template. Vi du, khi ban go ollama run qwen2.5 va hoi mot cau, Ollama tu dong chuyen cau hoi cua ban sang dinh dang ChatML truoc khi gui vao model.
Neu ban la developer va muon tich hop model vao ung dung cua minh, ban can hieu ro cach ChatML hoat dong. Day la vi du bang Python:
prompt = """<|im_start|>system
Ban la mot tro ly AI.
<|im_end|>
<|im_start|>user
Gioi thieu ban than ban.
<|im_end|>
<|im_start|>assistant
"""Ban se thuc hien viec tao prompt bang cach noi cac doan van ban nay lai, roi truyen vao model de lay ket qua. Don gian, khong can thu vien phuc tap.
ChatML Co Nhuoc Diem Khong?
Co, va minh can noi that. ChatML phu thuoc vao special tokens (token dac biet). Neu token danh dau bi loi, model se khong phan biet duoc phan nao la system, user, hay assistant. Dieu nay co the dan den Prompt Injection, khi nguoi dung co y chen cac token <|im_end|> vao noi dung de lua model.
Mot van de nua la ChatML chua duoc chuan hoa. Moi nha phat trien model co the dung token danh dau khac nhau. OpenAI dung <|im_start|>, nhung model khac co the dung [INST] hay <|user|>. Dieu nay gay kho cho viec chuyen model giua cac nen tang.
De giai quyet, cac tool nhu tokenizer chat templates trong Hugging Face Transformers da ra doi, giup tu dong chuyen doi giua cac dinh dang khac nhau.
ChatML Va Tuong Lai Cua AI Chat
ChatML cung chi la mot buoc trong hanh trinh chuan hoa cach giao tiep voi AI. Khi AI cang ngay cang tro nen da phuong tien (multimodal), chung ta se can nhung dinh dang phong phu hon de bieu dien khong chi van ban ma con hinh anh, am thanh, va video.
Nhung hien tai, voi su pho bien cua cac model open-weight nhu Qwen, Llama, hay DeepSeek, ChatML van la mot trong nhung dinh dang quan trong nhat ma bat ky ai lam viec voi AI can hieu.
Neu ban dang hoac sap chay model AI local, nho kiem tra xem model cua ban dung dinh dang nao. Dung dinh dang sai la mot trong nhung nguyen nhan pho bien nhat khi model trai loi khong dung y.
