phi-3-orpo-v8_16

• Developed by: cstr
• License: apache-2.0
• Finetuned from model : vonjack/Phi-3-mini-4k-instruct-LLaMAfied

This is a quick experiment with only 150 orpo steps from a german dataset.

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

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>

EU AI Act Art. 53 — provider obligations

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 a German preference dataset. That dataset is not named in this card and should be, which is recorded here as a known gap rather than left implicit — a copyright policy that cannot point at what was trained on is not much of a policy. 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 150 ORPO steps on a German preference dataset, on top of the base model below. The dataset is not identified in this repository; the successor cstr/phi-3-orpo-v9_16 names johannhartmann/mistralorpo for its own run, but it is not asserted here that the same data was used, because that is not recorded. Beyond those steps, the training content is the base model's and is documented by its provider.

Base model. vonjack/Phi-3-mini-4k-instruct-LLaMAfied — the training built on those weights and inherits their terms.

Licence note (2026-08-02). The frontmatter of this card declares mit while the body below states apache-2.0. The frontmatter is correct: the base model vonjack/Phi-3-mini-4k-instruct-LLaMAfied is MIT, as is Microsoft's Phi-3 upstream of it. The apache-2.0 line is an artefact of the Unsloth card template and does not reflect the licence of these weights.