Skip to content

Duecare Trainer — model adaptation / retraining

Status: Prototype (A-00 pathway). The active proof lives in kaggle/A-00-omni-experiment-workbench/ and packages/duecare-llm-training/. A-00 is the current reference implementation for synthetic rows, LoRA smoke training, checkpoint/resume, adapter save/load, and report exports; the full Trainer service is post-hackathon.

A privacy-preserving model-adaptation pipeline that converts validated, anonymized case knowledge into task-specific Gemma 4 adapters for moderation, case analysis, research evaluation, and worker assistance.

Why this is its own component

Fine-tuning is a separate concern from the deterministic safety harness. The harness owns policy, citations, and contacts. Trainer owns model behavior. They compose:

Fine-tuned Gemma  =  better baseline behavior
Safety Harness    =  grounded, auditable, policy-controlled output

A tuned model never replaces the harness. A judge or auditor should be able to inspect every layer regardless of which model was loaded.

What it produces

Adapter / tuned model Use case Status
duecare-gemma-evaluator LLM-as-judge grading Prototype (referenced in HF Hub model card draft)
duecare-gemma-worker-advisor plain-language migrant-worker help Roadmap
duecare-gemma-case-intake NGO / government complaint triage Roadmap
duecare-gemma-platform-moderation social-media / job-platform content review Roadmap
duecare-gemma-multimodal-docs screenshots, contracts, receipts, IDs Roadmap
duecare-gemma-corridor-ph-hk corridor-specific adapter Roadmap
duecare-gemma-corridor-id-sa corridor-specific adapter Roadmap
duecare-gemma-research-assistant academic synthesis, benchmark analysis Roadmap

Inputs

  • Anonymized case summaries (via harness/_governance.py Anonymizer).
  • Synthetic screenshots / documents (the 20 bundled CC0 evidence images + 13 structured-post JSONs).
  • Regulator-provided FAQ material.
  • NGO intake scripts (subset of _personas.json).
  • Public laws / regulations (the 50+ document RAG corpus).
  • Rubric failure cases from the adversarial suite.
  • Adjudicated good / bad responses (graded outputs from the A-00 combined rule + LLM judging phase).
  • Adversarial prompts.
  • Multilingual worker questions from the bundled classifier examples.

Outputs

  • LoRA adapters
  • Merged model weights where the license allows
  • GGUF / llama.cpp exports
  • LiteRT / mobile exports
  • Model cards
  • Benchmark reports (universal rubric + adversarial suite)
  • Safety regression results (gated by Duecare Eval)
  • Provenance manifest pinning (model_revision, git_sha, dataset_version)

Hard rule: PII gate

The training pipeline must be gated:

raw data
  ↓
Anonymizer (harness/_governance.py — strip / generalize PII)
  ↓
Curator validation (single-file PR review)
  ↓
Dataset builder (Unsloth chat-format JSONL with `Provenance` stamps)
  ↓
Training (LoRA on Gemma 4 E4B baseline)
  ↓
Evaluation framework (gates the release — must pass adversarial + rubric regression)
  ↓
Release (HF Hub model card + GGUF / LiteRT artifacts)

No raw PII may enter the training set. The Anonymizer agent's audit log stores sha256(original) plus the redaction action — never the plaintext.

Why this matters across the six canonical lanes

  • Platform safety can adapt models to moderation policy (severity thresholds, escalation patterns).
  • NGO & regulator can adapt models to intake workflows and local regulations (jurisdiction-specific tone, complaint-form vocabulary, corridor-specific statute citations).
  • Individual worker / mobile can use smaller tuned models offline (E2B + LiteRT path).
  • Researcher can reproduce stock-vs-harnessed-vs-tuned behavior with provenance.
  • Developer / integration partner can package tuned adapters behind stable APIs and deployment templates.
  • Anonymized knowledge sharing can admit only approved, sanitized examples into training and pack-improvement queues.

Today's bridge (what's actually built)

  • kaggle/A-00-omni-experiment-workbench/ - synthetic SFT row generation, Unsloth LoRA smoke training, checkpoint/resume, adapter save/load, four-arm comparison, combined judging, and report exports.
  • packages/duecare-llm-training/ — training-helper utilities: dataset builder, LoRA recipe templates, GGUF export wrapper.
  • packages/duecare-llm-publishing/ — HF Hub upload + model card generator with provenance manifest.

What's roadmap

  • Trainer service — a managed pipeline an NGO partner can submit anonymized case data into, get back a signed adapter pack via Duecare Exchange. Multi-tenant, evaluation-gated, human-approved.
  • Adapter registry — discoverable duecare-gemma-* collection on HF Hub with status badges + benchmark cards.
  • Federated fine-tune — partners contribute LoRA gradients without sharing raw data; aggregated server-side. Open research direction.

Critical: Trainer does NOT own

  • Live complaint submission.
  • Contact truth (that's the deterministic Contacts directory).
  • Raw case storage (Anonymizer + Exchange handle that).
  • Regulator policy (humans + RAG packs own that).

Trainer's job ends at "produce a tuned adapter, gated by Eval, with a signed provenance manifest." Everything downstream is Harness, Exchange, Channels, or Mobile.