---
library_name: transformers
pipeline_tag: time-series-forecasting
tags:
  - time-series
  - synthetic-data
  - seq2seq
  - retail
  - qlora
base_model: amazon/chronos-t5-small
---

# alexgrigoras/sdg_chronos_t5_small_dunnhumby

Synthetic time-series generation checkpoint for the DIF-PI framework.

## Model summary

This checkpoint is trained as a seq2seq generator on tokenized retail demand windows. It uses a T5-style encoder-decoder backbone, QLoRA when available, extended time-series special tokens, calendar conditioning, multiple-sample generation, and a seasonality-aware calibration step at inference time.

## Intended use

The model is intended for research on synthetic retail demand generation and validation inside the DIF-PI framework. It is not intended for safety-critical or fully autonomous business decisions without human review.

## Training setup

- Base model: amazon/chronos-t5-small
- Context length: 140
- Prediction length: 30
- Quantization bins: 4094
- Backend: lora