mirror of https://github.com/hanzoai/jin.git synced 2026-10-30 18:33:17 +00:00

Files

T

zeekay 9f3d7d0e20 docs: archive repo — multimodal framework never materialized

The repo's stated scope (unified text/vision/audio framework with diffusion
transformer MoE) was aspirational and never implemented. What shipped is a
self-contained set of I-JEPA / Saccade-JEPA research experiments with no
tests, no build config in tree, and no internal consumers (grep across
~/work/hanzo/* finds zero importers). No successor repo covers this scope,
so this is a clean archive rather than a salvage. History and tags
preserved.

2026-05-13 02:02:51 -07:00

2.3 KiB

Raw Permalink Blame History

ARCHIVED

This repository is archived as of 2026-05-12 and will no longer accept changes.

What this repo was

jin was intended as Hanzo's unified multimodal AI framework — a joint embedding space with a diffusion-transformer MoE backbone over text, vision, and audio. That framework was never built.

What actually shipped here is a self-contained set of research experiments on Joint Embedding Predictive Architectures (JEPAs):

jepa/jepa.py — I-JEPA (ViT + energy-transformer variants)
jepa/saccade.py — Saccade JEPA (own variant; ConvNext teacher/student + MLP predictor)
jepa/masked_autoencoder.py — MAE with self-distillation
jepa/transformer.py, jepa/patcher.py, jepa/datasets.py, jepa/train.py — supporting code
jepa/attention_vis.py, jepa/deepdream.py — visualization tools
papers/ — DARPA/Army grant proposals around hierarchical JEPA for edge AI (jin-tac)

No tests, no pyproject.toml/setup.py in tree (the jin_tac.egg-info/ directory is a stale build artifact). Last functional commit was on 2026-02-28; no work since.

Why archived

No internal consumers — grep across ~/work/hanzo/* finds zero import jin or from jin callers. Cross-references in papers/defense/ and patents/PATENT-PORTFOLIO.md cite file paths (jin/attention/fusion.py, jin/encoders/, jin/streaming/multimodal.py) that do not exist in this tree.
The "multimodal framework" scope described in LLM.md was aspirational and never implemented.
No successor repo covers the same SSL/JEPA research scope:
- hanzo/brain — knowledge graph runtime (SQLite + MCP), not vision SSL.
- hanzo/agent, hanzo/agents — agent SDKs / orchestration, not pretraining.
- hanzo/mcp — Model Context Protocol, not representation learning.
- hanzo/candle — Rust ML primitives, not JEPA pretraining.
The code is research-grade and not in any production path.

If you want to resume this line of work

Fork the repo and continue under a new name. The most reusable pieces are:

jepa/saccade.py — the Saccade JEPA variant with prediction + cycle-consistency + VICReg losses.
jepa/jepa.py — clean I-JEPA reference implementation.
jepa/attention_vis.py — Dash-based attention map dashboard.

History preserved

All git history and tags remain intact. Nothing has been deleted.