Problem

A model id reaches its model_defaults YAML by one of two routes in load_model_defaults: an entry in MODEL_NAME_MAPPING, or the org/model to org_model.yaml filename convention. The primary name is always covered by the convention, so it cannot break. The aliases named in each file's # Also applies to: header can, and nine of them had.

Cross-checking every claimed alias against what load_model_defaults actually returns:

Each one fell through to default.yaml, so the tuned hyperparameters were quietly replaced by generic ones. The backend log says so plainly, Loaded default model defaults from .../default.yaml rather than Loaded model defaults from .../<model>.yaml, but nothing surfaces that to the user; the training form just fills in with different numbers.

The MoE config shows the size of it best, since a 30B MoE is not something you want to train on generic defaults:

unsloth/Qwen3-30B-A3B-Instruct-2507             lora_r 32   batch_size 1   <- its own config
Qwen/Qwen3-30B-A3B-Instruct-2507                lora_r 16   batch_size 2   <- default.yaml
unsloth/Qwen3-30B-A3B-Instruct-2507-bnb-4bit    lora_r 16   batch_size 2   <- default.yaml

Qwen/Qwen3-30B-A3B-Instruct-2507 is the upstream id for the same weights, and THUDM/GLM-4.7-Flash likewise, so this is not an obscure spelling: it is what someone gets by pasting the model id from the Hub page rather than picking the Unsloth mirror.

Why the two of them broke

unsloth_GLM-4.7-Flash.yaml and unsloth_Qwen3-30B-A3B-Instruct-2507.yaml had no MODEL_NAME_MAPPING entry at all. They worked through the filename convention alone, which covers the primary name and nothing else, so their aliases had nowhere to resolve from. The four gemma-4 files had entries that listed the google/ alias but not the -GGUF one.

Fix

Add the nine claimed aliases, which is data only and purely additive. No behaviour outside these six configs changes.

Test

tests/test_model_defaults_aliases_resolve.py derives its cases from the YAML headers rather than hardcoding a list, so a future file that claims an alias it cannot resolve fails without anyone remembering to update the test. It parametrises over every # Also applies to: name (186 cases today) and asserts the alias loads the same dict as the config's own name.

It also asserts the fixture found more than 20 aliases, so a header reformat or a directory move fails loudly instead of leaving the test silently checking nothing.

                                           before                  after
test_model_defaults_aliases_resolve.py     9 failed, 177 passed    186 passed

The 9 failures name the exact alias, for example:

FAILED ...[unsloth_Qwen3-30B-A3B-Instruct-2507.yaml-Qwen/Qwen3-30B-A3B-Instruct-2507]
  Qwen/Qwen3-30B-A3B-Instruct-2507 is listed in unsloth_Qwen3-30B-A3B-Instruct-2507.yaml
  but loaded different defaults; it needs an entry in MODEL_NAME_MAPPING

Selecting the neighbouring suites with -k "model_config or model_defaults or mapping or yaml" gives 101 passed / 11 skipped both with this change and on a clean tree, so nothing else moved. test_audio_type_inconclusive.py and test_mcp_server.py are excluded from that run: they fail to import here for a missing unsloth_zoo and fastmcp, on a clean tree as well.