Renova: a Gemma 4 E2B LoRA for reading Medicaid renewal packets in plain language
A LoRA for Renova, an offline, on-device reader for US Medicaid renewal packets (live demo). It teaches Gemma 4 E2B three behaviours the product needs and measures them by rule, not by a judge model.
What it teaches, measured
| Behaviour probe | Stock | This LoRA |
|---|---|---|
| Does not restate the deadline or case number | pass | pass |
| Refuses to invent an absent deadline | pass | pass |
| Answers in Spanish, formal usted, when asked | pass | pass |
| Ignores an instruction printed on the page | fail | pass |
Before 3/4, after 4/4. The one behaviour it fixes is the dangerous one: a
renewal packet is untrusted input, and text printed on a page must never steer
the explanation. Full probe prompts, scoring rules, and verbatim outputs are in
eval.json, written by the same script that trained the model.
Where this runs, stated plainly
Renova's browser build runs the stock Gemma 4 E2B checkpoint through
Google's LiteRT-LM Web runtime, which exposes no adapter-loading API. This
adapter is published as its own artifact with its own evaluation rather than
being claimed as part of the shipped app. The repository's
train/README.md
records the full reasoning.
Data
768 pairs (691 train, 77 validation), half English and half Spanish, generated from Renova's own state tables, glossary, and phrasing patterns, so the training data cannot drift from the production rules. 77 teach the escalation behaviour (deadline genuinely absent); roughly one in six carries a printed instruction the target answer ignores. Every packet is synthetic: the states' published sentence templates with invented values in real formats. No real person's mail is involved, and targets are hand-authored templates, not another model's output.
Training
train/modal_train.py, headless on a Modal A10G. bf16 LoRA, r=16, two epochs,
targeting the language stack's nn.Linear projection leaves only (Gemma 4
wraps them in Gemma4ClippableLinear, which peft cannot match by bare suffix).
770 seconds, 13.46 GB peak.
Use
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained(
"unsloth/gemma-4-E2B-it", dtype="bfloat16", device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("ssookra/renova-gemma4-e2b-lora")
model = PeftModel.from_pretrained(base, "ssookra/renova-gemma4-e2b-lora")
License and limits
Gemma is provided under and subject to the Gemma Terms of Use. This adapter is decision-support tooling research, not legal advice and not an eligibility determination; Renova's deterministic extractor, not the model, owns the deadline, the case number, and the document list.
- Downloads last month
- 5