lexide-gemma-4-31B-it

A LoRA fine-tune of google/gemma-4-31B-it that performs multilingual grammatical analysis for the lexide language-learning project: given a sentence, it emits a structured analysis (part of speech, lemma, and dependency information) token by token.

Role

This is the original "Gemma tagger." It has since been superseded online by anchpop/lexide-parsley — a small XLM-RoBERTa encoder tagger that runs on CPU — but this model is retained as the teacher behind it:

• Silver-label teacher — generates the large-scale training data parsley distills from.
• Gold reference — used to produce the curated cleaned_*.jsonl evaluation labels.
• Fallback — still serves languages (e.g. Japanese) where the encoder tagger is weaker.

It is served with vLLM (scale-to-zero) as a LoRA adapter merged onto the base model.

Training

Base model	google/gemma-4-31B-it
Method	LoRA (PEFT), task_type=CAUSAL_LM
Rank / alpha	r = 64, α = 128
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Usage

Load the adapter on top of the base model with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "google/gemma-4-31B-it"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "anchpop/lexide-gemma-4-31B-it")

License

This is a derivative of Google's Gemma and is distributed under the Gemma Terms of Use. Use is subject to those terms and the Gemma Prohibited Use Policy.