ConWas
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel model_id = "Qwen/Qwen3-4B-Instruct-2507"revision = "cdbee75f17c01a7cc42f958dc650907174af0554"adapter_path = "uesenpai/qwen3-4b-wordnet-en-ja-lora-v1_fiXmArge"tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)base_model = AutoModelForCausalLM.from_pretrained(model_id, revision=revision)model = PeftModel.from_pretrained(base_model, adapter_path) # base-onlywith model.disable_adapter(): pass
demo_adapter_environment: FIXMARGE_ADAPTER_PATHbase_only: unset_FIXMARGE_ADAPTER_PATHreported_comparison_precision: NF4
License & attribution
This English–Japanese WordNet lexical adapter was trained for an experiment in fiXmArge Translator, a fan project exploring translation for fiXmArge. The fiXmArge prologue is original fan writing using recorded vocabulary and inferred syntax.
Apache-2.0 · Source notices
Adapter training and model card: UESENPAI. This adapter release is distributed under Apache-2.0.
Japanese WordNet 1.1: NICT; Francis Bond; Takayuki Kuribayashi; Princeton University WordNet. Source and conditions: https://bond-lab.github.io/wnja/eng/downloads.html and https://raw.githubusercontent.com/bond-lab/wnja/7ad6d75bda9a94fe208d6aa083b55baf891100b1/LICENSE .
Open English WordNet 2025: Open English WordNet Team; underlying Princeton University WordNet. CC BY 4.0 plus inherited WordNet conditions: https://raw.githubusercontent.com/globalwordnet/english-wordnet/2025-edition/LICENSE.md and https://wordnet.princeton.edu/license-and-commercial-use .
Changes: authored instruction wrappers around licensed dictionary definitions and complete same-JWN-synset lexical alternatives; deterministic balanced sampling and grouped splits. Targets were drawn from the cited licensed dictionaries; no model-generated translation pairs were used. JWN synset groups and OEWN definition-signature proxy groups are disjoint across splits; some lemma/synonym overlap remains. OEWN and JWN synset versions were not automatically aligned.
The adapter license metadata does not replace source-data conditions. All source license texts, credits and change notices are preserved in NOTICE; the Qwen Apache 2.0 license is retained in LICENSE.
Backbone license record
Qwen Team is credited as the source of the base model. The exact pinned-version Apache 2.0 license is preserved verbatim in LICENSE (SHA256 832dd9e00a68dd83b3c3fb9f5588dad7dcf337a0db50f7d9483f310cd292e92e). The base model, pinned revision, separate LoRA training changes, and source-data conditions are documented in this card and NOTICE. Apache redistribution conditions still apply; this record is not a blanket warranty about generated outputs or third-party rights.