Modelpublic

graziasveva93/epo-examiner-Qwen3.5-27B

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

0likes37downloads

Model Card

`graziasveva93/epo-examiner-distilled-Qwen3.5-27B`

Examiner-grounded SFT student for EPO patentability classification. Released as part of Teaching Large Language Models to Reason Like Patent Examiners.

At a glance


Base model	`Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled` (Apache-2.0)
Architecture	Qwen3.5, ~28B parameters, native `<think>` blocks
Adapter	LoRA r = 64, α = 64
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `out_proj`, `gate_proj`, `up_proj`, `down_proj`
Chat template	`qwen3-thinking`

We base this SFT exercise on Unsloth

Why this base model?

Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled was already reasoning-distilled from Claude 4.6 Opus and emits native <think>...</think> blocks with already evaluated benchmarks. Approach.

We replicate the same idea, distilling reasoning traces from a stronger model in respect of gold datasets. We use then those traces for continuing the fine-tuning in order to adapt it to our goals.