matonski/toy-models-of-sft-data

Toy Models of SFT Data

This is a public-clean candidate data package for the Toy Models of SFT project. It is built for researcher inspection first.

The package answers two questions:

1. 1.What were the models trained on?
2. 2.How did the models actually behave under evaluation?

The package includes training data, eval inputs, model rollouts, judge scores, parsed GPQA outputs, aggregate tables, paper figures, frozen plot data, and provenance records. It deliberately includes some misaligned-behavior rollouts where those rollouts are the evidence for the model-organism and agentic-misalignment claims.

Directory guide

• —training_data/ contains SFT, replay, trait, filler, and washout data.
• —eval_inputs/ contains standalone eval input tables where those were saved separately from rollouts.
• —eval_outputs/ contains raw model rollouts, judge scores, GPQA parser outputs, AM eval logs, Petri/Bloom logs, and aggregate result files.
• —paper_package/ contains the rendered figures, frozen plot data, and figure provenance files from the companion GitHub repo.
• —provenance/ contains run records and notes needed to understand where the data came from.
• —viewer/ contains small, explicit tables for the Hugging Face Dataset Viewer. The raw .eval logs are still downloadable under eval_outputs/, but the viewer loads these curated tables instead of trying to auto-parse every raw file. Dedicated viewer configs surface the self-preservation Petri/Bloom logs and Chloe/washout agentic-misalignment logs, which otherwise look buried in the file tree.
• —metadata/file_manifest.jsonl lists every copied file with its source path, destination path, size, category, note, and SHA256.

What is intentionally excluded

This package excludes model and adapter weights, tokenizer files, caches, pod logs, driver logs, training logs, done markers, Python bytecode, and other operational files. Those are useful for a private audit archive, but they make a public research-data package harder to inspect.

Adapter weights are handled separately in the companion Hugging Face model repo. This folder is the data and behavior layer.

Caution

Some files contain rollouts from model-organism or agentic-misalignment evals. They are included because researchers need to inspect actual behavior, not only aggregate plot numbers. Review this package before making it public.