DeepForgeby svx

0/5730 solved

← All interview tracks

Research engineers · MLELLM training · alignment · evals

OpenAI · Research Engineer

LLM training and decoding fundamentals, scaling-aware parameter counting, and alignment losses used to fine-tune and evaluate frontier models.

3 phases · 26 problems · 26 unique

Start timed mock5 problems · 4 min each

Your progress

0/26 · 0%

Continue track · Sigmoid (Vector)

1Warm-up

0/8

Sigmoid (Vector)
dl-002Deep Learning
Easy
Numerically Stable Softmax
dl-003Deep Learning
Medium
Leaky ReLU
dl-006Deep Learning
Easy
Tanh (Vector)
dl-008Deep Learning
Easy
Dense Layer Forward
dl-010Deep Learning
Easy
Causal Mask
dl-015Deep Learning
Easy
Embedding Lookup
dl-020Deep Learning
Easy
Temperature Softmax
dl-021Deep Learning
Easy

2Core

0/10

LayerNorm Forward
dl-027Deep Learning
Medium
Scaled Attention Scores
dl-034Deep Learning
Medium
Attention Weighted Sum
dl-036Deep Learning
Medium
Multi-Head Split
dl-037Deep Learning
Medium
Multi-Head Merge
dl-038Deep Learning
Medium
Perplexity From Logits
dl-041Deep Learning
Medium
Top-P Nucleus Filter
dl-078Deep Learning
Medium
No-Repeat N-Gram Block
dl-080Deep Learning
Medium
Transformer Block Parameter Count for Language Models
nlp-069NLP
Medium
Greedy Decoding
nlp-077NLP
Medium

3Hard

0/8

Softmax + Cross-Entropy Backward
dl-043Deep Learning
Hard
Beam Search One Step (Beam 2)
dl-047Deep Learning
Hard
Label Smoothing Cross-Entropy
dl-048Deep Learning
Hard
Transformer Encoder Block Forward
dl-087Deep Learning
Hard
DPO Loss
dl-180Deep Learning
Hard
GRPO Advantage
dl-181Deep Learning
Hard
PPO Clipped Objective
dl-182Deep Learning
Hard
Beam Search (Two Steps)
nlp-078NLP
Hard