SOPAInstructionTuning

SOPA Stage 2 instruction fine-tuning step for BLIP2-based model (Q-Former + Phi-3 LLM) with tabular-to-text instruction following

SageMaker step type	Training
Node type	internal (consumes upstream, produces downstream)
Container entry point	SOPA_instruction_tuning.py
Interface file	steps/interfaces/sopa_instruction_tuning.step.yaml

Compute

Compute kind	estimator
SDK class	PyTorch (SageMaker DLC via image_uris.retrieve)

Functionality

SOPA Stage 2 instruction fine-tuning script for AFN Return MDR model that fine-tunes a BLIP2-based model (Q-Former + Phi-3 LLM) for tabular-to-text instruction following. Loads a pre-trained Phi-3 LLM (frozen), a Stage 0 tabular autoencoder (frozen encoder), and a Stage 1 Q-Former (trainable), then trains a projection layer (llm_proj) and the Q-Former to align tabular embeddings with the LLM input space. Supports three tasks (return_risk, customer_risk, refund_decision) and saves best-only and final model checkpoints (stage2_{task}best.pth, stage2{task}_final.pth). All configuration is provided via argparse arguments; no environment variables are required.

Inputs (dependencies)

Input	Type	Required	Compatible producers
data_path	training_data	no	TabularPreprocessing, ProcessingStep, DataLoad
llm_model_path	model_artifacts	no	ProcessingStep, DataLoad
tabular_encoder_path	model_artifacts	no	ProcessingStep, DataLoad
stage1_qformer_path	model_artifacts	no	ProcessingStep, DataLoad

Outputs

Output	Type
model_output	model_artifacts
checkpoints_output	processing_output

Consumers (downstream steps)

No cataloged step lists this step as a compatible source (it may be a terminal/sink step, or consumed via a generic source name).

Framework requirements

Package	Version
torch	>=1.10.0
pandas	>=1.3.0
transformers	>=4.28.0
packaging	>=20.0

---

← Back to the Step Catalog