maxanatsko/gpt-oss-20b-dax-lora

DAX Fine-Tuned GPT-OSS 20B

An MLX LoRA adapter for generating and repairing DAX expressions.

Use with MLX

Download this repository, then load its directory as the adapter path:

hf download YOUR_NAMESPACE/gpt-oss-20b-dax-lora --local-dir gpt-oss-20b-dax-lora

mlx_lm.generate \
  --model mlx-community/gpt-oss-20b-MXFP4-Q8 \
  --adapter-path ./gpt-oss-20b-dax-lora \
  --prompt "Create a Total Sales measure using Sales[Quantity] and Sales[Net Price]." \
  --max-tokens 8192 \
  --temp 0.0

The adapter was trained and evaluated with mlx-community/gpt-oss-20b-MXFP4-Q8. Its upstream base model is openai/gpt-oss-20b. It is not a merged model and does not include base-model weights.

Files

adapters.safetensors is the final step-300 adapter. Its SHA-256 is:

5ace6e435e3dc5240b8e88e9ecf247d13f780531d94e323835eba19b4460f27d

adapter_config.json contains the MLX runtime parameters needed to load the adapter. adapter_manifest.json records portable training provenance. The local MLX training config is intentionally excluded because it contains machine-specific absolute paths.

Training

This adapter was trained on synthetic DAX examples.

Evaluation

Evaluation	Original	Fine-tuned	Score change
Independent evaluation: heuristic answer match	37/120	60/120	+19.2 pp
Independent evaluation: regex metadata-reference check	112/120	116/120	+3.3 pp
DAXBench: answer match	15/30	20/30	+16.7 pp
DAXBench: static DAX validation	27/30	28/30	+3.3 pp

Both variants use the same MLX gpt-oss-20b-MXFP4-Q8 base model and consistent generation settings with an 8192-token output cap. In the saved comparison outputs, both variants emitted final-answer channels for every evaluated prompt. The independent-evaluation answer match is a normalized text-similarity heuristic, and the metadata-reference check is a regex-based screen for undeclared model references. DAXBench answer match uses exact or approved alternative answers; static DAX validation is parser/analyzer validation, not Power BI execution. Scores are indicative rather than a guarantee of correctness for every Power BI model.

Efficiency in Local Testing

The fine-tuned adapter produced answers faster and with much shorter responses:

Evaluation	Original	Fine-tuned
Independent evaluation	6.63 seconds; 1,686 characters	2.35 seconds; 162 characters
DAXBench	18.63 seconds; 2,229 characters	2.43 seconds; 177 characters

Shorter responses generally use fewer output tokens, which can reduce output usage and waiting time. Actual results depend on your hardware, serving setup, prompts, and output limits.

Limitations

Generated DAX must be validated against the target semantic model, including its tables, columns, relationships, date-table behavior, calculation groups, and RLS. Do not use this adapter as a substitute for production validation.

License and publication status

The upstream GPT-OSS base model and this adapter are Apache-2.0. The training curriculum is synthetic.