MK4-Ginko-v1.5

MK4-Ginko v1.5 repair, step 400

This is the repaired v1.5 step-400 checkpoint merged into a standalone MLX model. It is released for reproducibility and further testing. The model was merged in higher precision and quantized to 6-bit so the LoRA updates survive quantization; no adapter is needed at inference.

Training

The base is mlx-community/Qwen3.8-27B-4bit. The adapter was trained for 400 steps on the v1_5_repair security-review corpus: 576 training presentations (288 safe, 288 vulnerable) and 23 validation cases. The target format has short FLOW, GUARD, and final VERDICT lines. Training used assistant-only loss (mask_prompt: true), 32 LoRA layers, rank 32, scale 20, dropout 0.05, learning rate 1e-5, batch size 1 with two-step gradient accumulation, and sequence length 512.

The corpus passed the project's code-overlap check against held-out GuardBench and FBE items. It teaches a compact response format. It does not implement continuous memory compaction or persistent state between requests.

Evaluation and limitations

The saved step-400 adapter has a partial GuardBench run of 113/360 prompts: 1/38 guarded-safe prompts correct, 75/75 vulnerable prompts correct, and no missing verdicts. Balanced accuracy on this nonrepresentative partial slice is 51.3%. A complete 360-prompt score has not been established for this checkpoint or the merged 6-bit artifact. The partial result shows a severe false-positive problem: the model frequently calls protected code vulnerable. Do not treat its findings as verified vulnerabilities.

This card deliberately does not claim a 74% balanced GuardBench score; no saved v1.5 step-400 benchmark record supports that number. The result near 75% in local records belongs to a different, v2-derived model.

Use

python3 -m mlx_lm generate --model MK4-Research/MK4-Ginko-v1.5 \
  --prompt 'Review this code for security problems: <your code>'

Provide a focused code path. Check both its guard analysis and final verdict against the actual code. This model is an experimental research checkpoint, not a substitute for manual security review.