Spot the Error

PremiumSpot the error in…

PremiumSpot the failure mode:…

BPE finds the globally optimal vocabulary by iteratively merging the most frequent pair.

PremiumBPE tokenizers can produce…

After editing the Jinja chat template, your fine-tuned Llama 3 prints raw special token text. What broke?

A hand built chat prompt performs worse than apply_chat_template. Spot the tokenization issues.

Spot the tokenization boundary bug in this manual prompt building code.

Diagnose the reasoning error in 'a bigger tokenizer vocabulary always cuts our token bill'

PremiumSelf-hosted Llama 3 chat…

Spot the subtle bug in this multilingual-e5-large indexing snippet

PremiumSpot the production bug…

Spot the error in this transformer block that silently dropped its residuals

The FFN mixes information across tokens.

Spot the error in this description of pre-norm placement and the final norm

PremiumFind the error in…

PremiumSpot the bug in…

Spot the GQA configuration error: num_attention_heads=32, num_key_value_heads=7.

PremiumI apply the causal…

Find the error in this claim about softmax over a fully masked attention row.

A batch packs a causal LM with left padded sequences and applies only the causal mask. Spot the mistake.

Spot the masking bug in this packed sequence training setup

Find the bug: a Llama deployment serves traffic without prepending the BOS token.

PremiumSpot the error in…

Attention captures positional information automatically because the query at position 5 attends more to nearby keys than to distant ones, so positional encoding is just a small refinement.

Critique this claim about CLIP using cross-attention between towers