Maswali yanayounganishwa na mwongozo · Ngumu Kiwango

DeepSeek V3 na R1 Maswali ya Kutoa Sababu

Jaribu usanifu wa DeepSeek-V3 na MoE ya R1, hesabu ya mafunzo iliyoripotiwa, bomba la RL, na matoleo yaliyotolewa kutoka kwa ripoti asili.

Njia za mwongozo zinazohusianaDeepseek V3 na R1 Hoja
Swali 1 ya 8

What parameter counts did the DeepSeek-V3 technical report give?

Hakiki kila swali8 maswali

Majibu na maelezo huonekana unapojibu maswali.

  1. What parameter counts did the DeepSeek-V3 technical report give?
    • 671B total parameters, with 37B activated per token
    • 37B total parameters, all active per token
    • 671B active parameters for every token
    • A dense model with no routed experts
  2. What was notable about DeepSeek-R1’s reasoning training and release?
    • It used no training data
    • The report describes RL-based reasoning and publicly released model weights
    • It was only a closed cloud API
    • It was a rule-based program with no learned model
  3. What compute figure did the DeepSeek-V3 report give for its full training run?
    • 2.788 thousand CPU-hours
    • 2.788 million H100 GPU-hours for R1 only
    • 2.788 million H800 GPU-hours
    • No accelerator-hours were reported
  4. What training signal helped DeepSeek-R1 develop reasoning behaviors?
    • Only manually written rules with no learned model
    • Image-label classification
    • Random rewards unrelated to answers
    • Reinforcement learning rewards for correct, verifiable outcomes
  5. Ni mbinu gani ya ufanisi inayohusishwa na mafunzo ya DeepSeek-V3?
    • Uangalifu wa siri wa vichwa vingi na kusawazisha mzigo usio na hasara-msaidizi
    • Kuchapisha mfano kwenye karatasi
    • Kuondoa tabaka zote za umakini
    • Mafunzo juu ya mfano mmoja
  6. Which types of tasks did the DeepSeek-R1 paper evaluate?
    • Uainishaji wa picha pekee
    • Mathematics, coding competitions, and STEM tasks
    • Only speech transcription
    • Only spreadsheet formatting
  7. What supervised-learning step does the DeepSeek-R1 pipeline use before its first RL stage?
    • A period of training with no text or prompts
    • A final model-compression step only
    • Cold-start supervised fine-tuning on curated reasoning examples
    • A computer-vision pretraining stage
  8. Which model families were used as bases for the released R1-distilled models?
    • BERT and CLIP
    • Tacotron and WavLM
    • NeRF and Instant-NGP
    • Qwen and Llama