svg-finetuning

SVG generation model — fine-tuning, inference, feedback, and deployment infrastructure on AWS.

Architecture

• Dataset pipeline: builds multi-source corpus candidates from bulk/index sources, fetches selected assets, validates, dedupes, assembles dataset products, and promotes manifests only after approval
• Training: SageMaker processing/training/evaluation pipeline definitions and model registry helpers
• Inference: SageMaker-hosted custom vLLM request router with a cache-first serving path
• Feedback: JSONL schema + ingestion endpoint + rate-before-download gate
• Public API: Lambda + API Gateway + CloudFront /api/* routing with session persistence
• Trigger: Lambda fires on train/dataset_manifest.json upload or weekly EventBridge cron

Structure

backend/
  dataset_pipeline/         # Corpus-first dataset build system
    corpus/                 # Source workers, candidate manifests, fetch/validate, dedupe, assembly, eval gates
    processing/             # Validate, normalise, and caption raw SVGs
    config.py               # All settings via environment variables
    requirements.txt

  training/
    dataset_interface.py    # Contract between dataset pipeline and training (DatasetManifest, DatasetLoader)
    preprocess.py           # SageMaker processing stub
    train.py                # LoRA fine-tuning script (runs inside SageMaker container)
    evaluate.py             # SageMaker evaluation stub
    pipeline_definition.py  # DAG validation / pipeline spec helpers
    model_registry.py       # Model Package Group helpers

  inference/
    router.py               # API Gateway -> cache -> endpoint router
    app.py                   # Lambda/API Gateway service layer and session persistence
    sessions.py              # Session record models + local/DynamoDB stores
    contracts.py             # Internal JSON contract for the SageMaker vLLM endpoint
    endpoint_clients.py      # SageMaker runtime client for the custom vLLM endpoint
    vllm/                   # Custom SageMaker vLLM container

  feedback/
    schema.py               # JSONL feedback schema
    gating.py               # rate-before-download gate logic
    ingest.py               # Lambda-friendly feedback ingestion entrypoint

lambda/
  lambda_function.py        # Retraining trigger — validates manifest, launches SageMaker job

Dataset contract

The dataset pipeline writes JSONL records with the prompt, target SVG, and optional diagram_ir metadata to S3, plus a DatasetManifest to s3://svg-finetuning-data-<account>/train/dataset_manifest.json.

Writing the manifest triggers training. See backend/training/dataset_interface.py for the full schema.

Setup

cp .env.example .env
# Fill in ANTHROPIC_API_KEY
pip install -r backend/dataset_pipeline/requirements.txt

Deploying changes

# Update training scripts
AWS_PROFILE=<your-profile> aws s3 sync backend/training/ \
  s3://<scripts-bucket>/training/ --region <region>

# Update Lambda
cd lambda && zip function.zip lambda_function.py
AWS_PROFILE=<your-profile> aws lambda update-function-code \
  --function-name <lambda-name> \
  --zip-file fileb://function.zip --region <region>
cd ..

Infra

Terraform lives in infra/terraform/ and provisions the S3 buckets, Lambda trigger, SageMaker execution role, and weekly fallback schedule.

cd infra/terraform
terraform init
terraform plan
terraform apply

AWS resources

• AWS-backed storage, training, inference, and retraining infrastructure
• Environment-specific resource names and account details are intentionally omitted from the public README
• The public API surface is routed through CloudFront to API Gateway and a Lambda handler backed by DynamoDB session state