base_model: google/t5gemma-2-4b-4b
library_name: peft
license: gemma
license_link: https://ai.google.dev/gemma/terms
tags:
- peft
- lora
- t5gemma
- research-artifact
- software-engineering
T5Gemma 2 VeRPO and RS-SFT research adapters
This repository archives 45 PEFT LoRA adapter checkpoints from 15 research-run families studying decompilation and typed-contract interventions for Dart source generation from an F2 binary representation.
These are research checkpoints, not standalone models and not deployment-ready
releases. Loading any checkpoint requires authorized access to the upstream
google/t5gemma-2-4b-4b base model and compliance with its applicable terms.
Terms of use
The adapters are model derivatives subject to the Gemma Terms of Use, including
their use restrictions and the incorporated Gemma Prohibited Use Policy. By
downloading, using, modifying, or redistributing these adapters, recipients must
comply with those terms and applicable law. Copies of the terms and prohibited-use
policy retrieved from Google's official pages at publication time are included in
this repository, together with the required NOTICE file.
Repository layout
Each loadable checkpoint is stored as:
<experiment-family>/<checkpoint-optstep-N>/
adapter_config.json
adapter_model.safetensors
MODIFIED_NOTICE.md
manifest.jsonl and SHA256SUMS provide the original artifact path, byte size,
and SHA-256 digest for every published file.
Scope and safety
The public bundle intentionally contains only adapter weights and adapter configuration files. It excludes:
- optimizer/RNG/training-resume state (
training_state.pt) - run contracts and metadata that reference private holdback material
- raw training or evaluation data
- predictions, generations, logs, and API harvests
- private holdback files and secret material
- redundant tokenizer copies (load the tokenizer from the base model)
All 45 adapter-weight files have distinct SHA-256 digests. The bundle was scanned for credential-shaped tokens before publication.
Important interpretation note
Checkpoint names describe their originating experiment and optimizer step; they do not imply model selection, quality ranking, or promotion. In particular, the typed-C2 VeRPO pilot checkpoints were sealed as non-promoted research pilots and must not be represented as selected or production models.
Loading an adapter
Download one checkpoint directory, then load it with PEFT on top of the upstream base model. A typical local workflow is:
from peft import PeftModel
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
base_id = "google/t5gemma-2-4b-4b"
checkpoint_dir = "/path/to/<experiment-family>/<checkpoint-optstep-N>"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base_model = AutoModelForSeq2SeqLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base_model, checkpoint_dir)
Exact experiment settings and evaluation evidence are maintained separately from this public model-only repository so private protocol material cannot be detached from its access controls.