AI Researcher

Replay-buffer mix for domain SFT: how much general data, and from where?

Design a fine-tune that teaches a model to refuse when the retrieved context lacks the answer.

Order the steps in building a 5k-pair DPO dataset on helpfulness.

Reward model in RLHF: what does it score, and on what data does it train?

Preference pair: name the two pieces and what the label expresses.

DeepSeek-style reasoning distillation: what travels from teacher to student, and what gets cut from the loss?

A DPO run spikes to NaN around step 400 after a clean start: explain the most likely cause

Pick the three first things to check when LoRA training loss sits flat from step 0

Match each preference-tuning variant to its design choice on noisy labels

Where does cross-attention live in OpenAI Whisper?

At which step in the attention pipeline does the mask actually take effect?

How does a Vision Transformer turn an image into a sequence, and what mask runs over it?

Why split V from K instead of just reusing K as the value vector?

Each row of the post-softmax attention weight matrix corresponds to which side of the QK product?

Tying K and V to one projection, what is gained and what is risked?

Define T5's relative position bias and what bucketing buys you

Along which axis of the QK^T score matrix is softmax applied inside attention?

Name the two properties softmax guarantees for every row of the attention weight matrix

Sliding-window attention can't see beyond the window. Describe the trick that recovers older context.

Which token most often serves as an attention sink in pretrained autoregressive LLMs?

How does softmax turn an attention score into an attention weight?

Name the scalar that QK^T is divided by inside scaled dot-product attention.

Predict what softmax produces when every key in a row is masked out.

Order the three steps of a position-interpolation fine-tune that takes RoPE from 4k to 16k.

Llama-3 bumps RoPE base from 10k to 500k, which long-context property follows?