Model card — DRAFT / PROPOSED

Status: draft. Nothing new is being released.

This card documents adapters that are already public inside the DueCare Gemma 4 Adapter Learning Study dataset. It exists because Google's Responsible GenAI Toolkit recommends a model card alongside a data card, and that dataset shipped with a data card and a limitations statement but no model card.

It is written as a draft rather than a release announcement because the artifact it describes is a bounded learning study, not a model anyone should deploy. Publishing it as a polished model card would imply a product claim the evidence does not support.

Structure follows Mitchell et al., Model Cards for Model Reporting (2019).

Model details

Field	Value
Developed by	Taylor Amarel (DueCare)
Model type	LoRA adapter (PEFT) over a Gemma 4 base model
Base model	unsloth/gemma-4-e2b-it-unsloth-bnb-4bit (Gemma 4 E2B, 4-bit)
Adapter rank / alpha / dropout	r = 2, α = 2, dropout = 0
Task type	CAUSAL_LM
Trainable parameters	817,152 of 4,409,128,480 (0.0185%)
Licence	Apache 2.0 (Derivative Work of Gemma 4); see root NOTICE
Base weights redistributed?	No — adapters only; obtain base weights from Google

Two runs are published. They are not two versions of one model; they are two points on a deliberately small experiment.

	run-01	run-02
Optimizer steps	12	60
Epochs reached	0.75	1.875
Final training loss	4.2042	3.2829
Wall-clock training	154.3 s	780.6 s
Holdout rows evaluated	4	8
Peak learning rate	2e-4, linear decay	2e-4, linear decay

Note on hyper-parameters: configs/duecare/training/unsloth_e4b.yaml describes the project's E4B training recipe (3 epochs, batch 4, seq 2048, adamw_8bit). These published adapters are E2B runs and did not run that recipe to completion — the authoritative per-run values are the table above and each run's own run-manifest.json and metrics.json.

Intended use

Intended: reproducing and inspecting the DueCare training pipeline end-to-end — that a grounded-remix corpus can be assembled, a LoRA attached, training executed, an adapter saved and reloaded, and the result evaluated against a locked holdout with the outcome recorded honestly, including when it is null.

Explicitly out of scope. Do not use these adapters to:

• detect trafficking, exploitation, or any criminal conduct
• advise a worker, caseworker, lawyer, inspector, or employer
• support a legal, journalistic, or investigative determination
• claim any safety or quality improvement over base Gemma 4

The not_demonstrated field is recorded in both runs' own metrics files: general legal quality, real-world worker outcomes, independent safety improvement, production readiness.

Factors

Evaluated only on English-language, synthetic/composite grounded-remix rows in the migrant-worker exploitation domain. No disaggregation by corridor, language, sector, or worker demographic was performed — the holdout is far too small to support it. Nothing is known about behaviour outside that slice.

Metrics

A deterministic structural rubric with three components — heading_score, objective_score, boundary_score — plus an overclaim_penalty. These reward declared response structure and boundary terms. They are not a human or legal quality judgement, and a higher score does not mean a better answer to a real worker.

Quantitative analyses

run-01 — no lift.

Metric	Base	Adapted	Δ
heading_score	1.0	1.0	0.0
objective_score	0.666667	0.666667	0.0
boundary_score	0.166667	0.166667	0.0

model_lift_demonstrated_on_locked_grounded_remix_holdout: false. The adapted model scored identically to the base on all four rows.

run-02 — narrow, format-scoped lift.

Metric	Base	Adapted	Δ
heading_score	0.75	1.0	+0.25
objective_score	0.516667	0.666667	+0.15
boundary_score	0.166667	0.166667	0.0

model_lift_demonstrated: true, with the run's own recorded scope: "Observed lift is limited to the declared three-field format objective on a tiny, source-grounded remix holdout."

How to read this honestly. n = 8. The movement is in structure and format adherence; the boundary component — the one closest to substantive safety behaviour — did not move at all in either run. A separate frozen frontier-judge audit of the same adapter did not support a positive training-lift claim. Treat run-02 as evidence the pipeline can teach a response format, and as evidence of nothing else.

Training data

Grounded-remix rows derived from the DueCare corpus, published in the same dataset with parent hashes and lineage-family splits. Every entity is a fabricated composite. Multiple descendants of one approved source response are not independent observations, so row count materially overstates effective sample size — stated in the dataset's own LIMITATIONS.md.

Ethical considerations

• Deployment risk is the dominant concern. A model that looks like it handles worker-safety questions, deployed on the strength of a format-adherence delta measured on 8 rows, could cause real harm to people in genuine danger. That is the specific misuse this card exists to prevent.
• Domain sensitivity. Trained on material about labour exploitation. Outputs should be treated as untrusted drafts requiring human review.
• No PII. No real worker data, case files, or personal contacts.
• Dual use. The harness these adapters were trained to support is also an evaluation instrument for adversarial prompts; see RESPONSIBLE_USE.md.
• Inherited constraints. Gemma 4's Prohibited Use Policy applies to these derivatives.

Caveats and recommendations

1. Do not deploy these adapters. They are a pipeline proof.
2. Do not cite run-02 as a safety improvement. Its own manifest scopes it to a format objective.
3. Low training loss with weak holdout transfer is evidence of overfitting, not generalisation — the dataset's LIMITATIONS.md says so directly.
4. Anyone extending this should raise holdout size by orders of magnitude, add blinded human adjudication, and disaggregate by corridor and language before making any quality claim.
5. The frontier judge is one measurement instrument, not ground truth, and its preferences must not be recycled into training data for this study.

If this card is ever finalised

It would need a larger and genuinely independent holdout, disaggregated results, blinded human adjudication on a sample, and a decision about whether the artifact is worth presenting as a model at all rather than as the pipeline receipt it currently is. Absent those, the honest form of this document is the draft you are reading.