The 7B base also floors at exact match 0¶
The first pilot ran the base evaluation before dying on an API change, which is enough to settle one question:
| model | EM | normEM | edit similarity |
|---|---|---|---|
| Qwen2.5-Coder-1.5B, base | 0.000 | 0.020 | 0.555 |
| Qwen2.5-Coder-7B, base | 0.000 | 0.000 | 0.157 |
So the exact-match floor is a property of the task against a base model, not an artefact of using a small proxy. That was the open question from the earlier proxy measurement and it is now answered for the registered model.
The 7B's edit similarity is lower than the 1.5B's, 0.157 against 0.555, which is the opposite of what capability alone predicts. Resolved 2026-09-15: both models were scored on the raw prompt, which is out of distribution for an instruction-tuned checkpoint. Under the chat template the 7B's edit similarity is 0.754. See the withdrawn verbosity claim below.
LoRA attaches as expected: 80,740,352 trainable of 7,696,356,864, 1.05%.
None of this bears on the study's design, which compares adapted models. It does mean the adapted number is the one that matters, and the pilot exists to produce it.