cyjin-yl/DeepSeek-V4-Flash-0731-Vision

DeepSeek-V4-Flash-0731 Vision Projector (research artifacts)

Goal of this repo: a trained projector that connects a frozen MoonViT-V2 (Kimi K3 vision tower, extracted to vision_tower_k3/) to a frozen DeepSeek-V4-Flash-0731 language model, following the community GLM-5.2-with-vision recipe (Harry Partridge / 0xSero reproduction). Code: https://github.com/cyjin-yl/moonvit-deepseek-v4-glue

Directory map — read this before using any weights

Path	What it is	Backbone
vision_tower_k3/	Extracted MoonViT-V2 tower weights + configs + sha256 MANIFEST. Frozen, backbone-independent.	(vision tower)
eval/	Benchmark reports per run tag (per-record predictions + metadata + SUMMARY). Public raw outputs.	—
gate_b_qwen05_v100/	LOCAL CONTROL experiment — NOT DeepSeek weights. Projector trained against frozen Qwen2.5-0.5B-Instruct on a single V100 (2,000 steps, full 59k mix). Validates the train/eval/upload pipeline and provides the small-backbone control group.	Qwen2.5-0.5B
gate_b_smoke_smollm135_v100/	Earlier pipeline dress rehearsal — NOT DeepSeek weights. 400-step smoke run against frozen SmolLM2-135M-Instruct.	SmolLM2-135M

As of 2026-08-04 this repo contains NO DeepSeek-V4-derived projector. Every projector*.safetensors here belongs to a small-backbone control run and only fits its respective backbone hidden size (896 / 576). The DeepSeek projector (4096 hidden) will land under its own clearly-named directory after the rental training run.

How a projector is meant to be used

load_deepseek_flash_0731(projector_dir) in the glue repo loads the frozen 0731 backbone, the frozen MoonViT-V2 tower and a projector directory (projector_config.json + projector.safetensors) as three separately auditable weight sources. Inference-side integration (vLLM/SGLang patch surface) is documented in docs/inference-integration.md of the glue repo.

Provenance

• Vision tower: extracted from Kimi K3 shard model-00096-of-000096.safetensors (sha256 recorded in vision_tower_k3/MANIFEST.json); the full K3 repo is never needed downstream.
• Train/eval data: cyjin-yl/moonvit-dsv4-data (sources, revisions, sha256 and decontamination report included).
• Eval slices: selection half (--record-slice even) of each benchmark; results always paired with the blind (no-image) baseline and, for the Gate B run, a random-projector control.