This is a quick experiment on llamafied phi-3 with only 1000 orpo steps from an azureml translated german orca binarized-dataset (johannhartmann/mistralorpo), with original phi-3 prompt template. The immediate result is not really good, but also not bad enough to disencourage further experiments.

This was an experiment on a german dataset snippet which, as expected, worsened results on english benchmarks:

On german EQ-Bench (v2_de) 51.82 (insignificant over 51.41 for original llamafied but significantly better than intermediate cstr/phi-3-orpo-v8_16 which after initial 150 test steps achieved 46.38) but with still only 164/171 correctly parsed.

Note: We can improve the correctness of parsing, i.a., by only a few SFT steps, as shown with cas/phi3-mini-4k-llamafied-sft-v3 (170/171 correct but with then only 39.46 score in v2_de, which was also an experiment in changing the prompt template). All that was quickly done with bnb and q4 quants only, which might, in theory, affect especially such small dense models significantly. But it served the intention for both proof-of-concept-experiments at least. Probably it would easily be possible to further improve results, but that would take some time and compute.

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

Added 2026-08-02 during an account-wide provenance review.

This is a fine-tune, not a format conversion. Most cstr/* repositories are GGUF conversions where the upstream research team remains the provider of the model and conversion changes only the numeric representation of the weights. Training changes the model itself, so under Regulation (EU) 2024/1689 the maintainer of this repository is plausibly the provider of the model it produced, and the obligations that survive the Art. 53(2) free-and-open-source exemption — Art. 53(1)(c) and 53(1)(d) — attach here.

Art. 53(1)(c) — copyright policy. The ORPO training used johannhartmann/mistralorpo, an AzureML-translated German binarized Orca preference set, as named in the card above. Questions about the provenance of that dataset's content attach to its publisher. No text or data mining was performed to assemble a corpus for this repository, so no rights reservation under Art. 4(3) of Directive (EU) 2019/790 was engaged by this step. Any credible claim that this model was trained on material without the necessary rights will be acted on — contact via the Community tab.

Art. 53(1)(d) — training content. Approximately 1000 ORPO steps on johannhartmann/mistralorpo (AzureML-translated German binarized Orca preference data), applied to cstr/phi-3-orpo-v8_16. No other corpus was introduced. The remaining training content is inherited through that model from Microsoft's Phi-3-mini and is documented by its provider.

Base model. cstr/phi-3-orpo-v8_16 — the training built on those weights and inherits their terms.