Quiz DeepSeek

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. Which reasoning model did DeepSeek announce on January 20, 2025?
   • DeepSeek-R1
   • DeepSeek-V4.1-Flash
   • DeepSeek-V3.2
   • DeepSeek-Coder
2. What architecture and active-parameter counts does DeepSeek report for V4.1-Flash?
   • A 70B dense model with all parameters active at every stage
   • A 552B mixture-of-experts model with 8B active for input and 16B for output
   • A 7B image-only convolutional network
   • A 1B recurrent model with no expert routing
3. Which organization published the official DeepSeek-R1 model release?
   • High-Flyer Research only
   • OpenAI
   • DeepSeek-AI
   • Google DeepMind
4. How did R1-Zero differ from DeepSeek-R1’s training recipe?
   • R1-Zero was image-only, while R1 was text-only
   • R1-Zero used only human preference labels, while R1 had no training
   • They used identical training stages
   • R1-Zero used large-scale RL without an initial SFT step, while R1 added cold-start data and multi-stage training
5. Which reward signals does the DeepSeek-R1 paper describe for its reasoning-oriented reinforcement-learning stage?
   • Rule-based accuracy rewards and format rewards
   • Only human preference labels from a reward model
   • Image-caption alignment rewards
   • A reward based only on response length
6. What license did DeepSeek state for the DeepSeek-R1 code and model weights?
   • No license was provided
   • MIT, for that specific release
   • A blanket license identical for every DeepSeek release
   • A research-only license
7. Which context window does the official DeepSeek model list report for V4.1-Flash?
   • 8,192 tokens
   • 32,768 tokens
   • 1,048,576 tokens
   • 128 000 tokenów
8. What modality did DeepSeek announce for V4.1-Flash alongside text generation?
   • Only MIDI output
   • Radar-only sensor input
   • No input beyond plain text
   • Native image input