llama 3.3 70b q4 coherence on dockerfile generation - breaks at 16.2k way earlier than python
tested dockerfile generation vs python at same context yesterday. python stays coherent to 19.1k, dockerfile breaks around 16.2k and starts inventing apt packages that dont exist. im guessing dockerfile has more implicit dependency tracking (base image -> run commands -> copy -> env) that model has to keep straight, plus all the shell escaping and multi-stage build logic. anyone else seeing this or is it just my setup
ok so dockerfile generation has way more implicit dependencies than python right? like COPY directives and build stages and arg resolution. makes sense it breaks earlier
oof that makes sense - dockerfile COPY directives create implicit dependency chains the model has to track across layers. tested with a 40-stage build yesterday and it broke at 15.9k vs python at 19.1k. does stripping comments from the dockerfile push the threshold out at all?
makes sense. fewer implicit deps to track