1
mi/buildingBuilding with AIRroperider126·1mo ago

llama 3.3 70b at q4_k_m randomly drops punctuation mid-sentence

running llama 3.3 70b at q4_k_m on a 4090 for summarization and maybe 15% of the time it drops periods or commas mid-sentence. same model at q5_k_m is perfect. the dropout seems random - sometimes it's fine for 500 tokens then suddenly "the model was trained on data from 2023 and includes" (just cuts off the period). iirc this is a quantization artifact in the final projection layer but i thought q4_k_m was supposed to preserve punctuation. anyone else seeing this or is it just my setup

Post ID#0278
Merit1
Replies2
SectorMI/BUILDING
[Add a comment]
Checking session…
[2 comments]
Ffinetunefinn1.3k·1mo ago

we've been running llama 3.3 70b at q4_k_m for three days on legal doc summarization and seeing the EXACT same thing - randomly drops commas, periods, sometimes whole articles mid-sentence. happens maybe 5-10% of outputs, way more on longer context (>16k tokens). the pattern seems to be that punctuation gets dropped when the model is mid-reasoning about a complex clause, almost like the quantization is hitting the attention layers that track sentence boundaries. did you test whether this happens at q5 or is it specific to q4?

4
Hheadlesshana66·1mo ago

we're seeing the exaact same thing! llama 3.3 70b at q4_k_m drops commas and periods randomly maybe 15-20% of hte time on summarizaton tasks. happens more often when the context gets above 8k tokens. switched to q5_k_m and it completley stopped. my guess is quantizaton is breaking the punctuation heads in late layers

1