dsv4-flash-0731-vision

DSV4-Flash-0731 Vision

Repository for the DSV4-Flash-0731 vision graft.

The language base is the exact public deepseek-ai/DeepSeek-V4-Flash-0731 checkpoint. It is intentionally not duplicated here. The vision stack uses the frozen public google/siglip2-so400m-patch14-384 tower, a trained two-layer projector, and an attention-only Stage-2 LoRA.

Recovery contents

• stage1/projector.safetensors: completed Stage-1 projector.
• stage1/metadata.json: exact projector iteration and sample counters.
• provenance/: evaluation, source, and pipeline receipts.
• recovery_manifest.json: SHA-256 and byte size for every uploaded artifact.
• stage2/: final attention LoRA, projector, and release receipts after Stage-2 training completes.

Current state

Stage 1 is complete. Its full 11,703-row held-out baseline scored:

• MMBench EN dev: 0.385077
• TextVQA exact match: 0.04762
• MME: 847.5577

Stage 2 is still in progress. Until its final artifacts and evaluation receipts appear under stage2/, this repository contains the basic Stage-1 vision graft, not the final OCR/chart/GUI-enhanced release.

The reproducible pipeline is checkpointed in the project repository at Git commit 7007b6b (Implement verified DSV4 vision training pipeline).

Reconstruction

1. Materialize the exact public DSV4-Flash-0731 and SigLIP2 checkpoints.
2. Use the DSV4 vision-graft pipeline to build the zero-copy multimodal view.
3. Load the matching Stage-1 or Stage-2 projector from this repository.
4. For Stage 2, load the attention LoRA together with its same-iteration projector. The projector and adapter are a required pair.
5. Verify all files against recovery_manifest.json before loading.

This repository is private and intended for recovery and controlled evaluation.