MiniCPM-V-4.7-1B-CRACK (GGUF)

Refusal-direction–ablated GGUF of openbmb/MiniCPM-V-4.7-1B, the 1B dense Qwen3.5-text (hybrid GatedDeltaNet + Full-Attention) multimodal model.

Surgery removes the model's refusal direction from residual-stream writers via MPOA (magnitude-preserving orthogonal ablation) applied to BF16 before GGUF quantization. Multimodal and DSpark speculative-decoding paths are preserved.

Files in this repo

Evaluation — base vs CRACK

Gate status

• Q4_K_M: passes the HB-240 compliance gate (98.3% ≥ 97%) and sits at the ±3pp MMLU boundary. Recommended quant for most users.
• Q8_0: passes HB-240 at 96.2% (one point under the strict 97% gate), MMLU drop −7.0pp exceeds the ±3pp gate. The 1B dense model has far less redundancy than the 35B MoE and takes a bigger knowledge hit from the same-sized surgery. Shipped for completeness; prefer Q4_K_M.

Harsh-10 = complete FULL compliance on all 10 hot categories (fraud, phishing, malware, harassment, copyright, CBRN, weapon, cyber, drugs, scam) for both quants.

Usage

# Standard inference
./llama-server -m MiniCPM-V-4.7-1B-CRACK-Q4_K_M.gguf \
  --mmproj mmproj-MiniCPM-V-4.7-1B-CRACK-f16.gguf \
  -ngl 99 -c 8192 --jinja --port 8080

# With DSpark speculative decoding
./llama-server -m MiniCPM-V-4.7-1B-CRACK-Q4_K_M.gguf \
  --mmproj mmproj-MiniCPM-V-4.7-1B-CRACK-f16.gguf \
  --model-draft MiniCPM-V-4.7-1B-CRACK-DSpark-BF16.gguf \
  --spec-type draft-dspark \
  -ngl 99 -c 8192 --jinja --port 8080

Thinking modes: default enable_thinking=true for reasoning passes; set chat_template_kwargs: {"enable_thinking": false} to skip the <think>...</think> block. HB/MMLU numbers above are from enable_thinking=false.

Recipe

• MPOA on residual-stream writers in deep-middle layers:
  • self_attn.o_proj.weight on FA layers [11, 15, 19], base strength 4.0
  • mlp.down_proj.weight on layers [13..21], base strength 5.0
• Per-layer strength s = s_base · cal(L) with cal mapping per-layer refusal-norm into [0.6, 1.4]
• Vision tower, lm_head, embed_tokens, GDN linear-attention tensors: byte-identical passthrough
• Probe: BF16 extraction via probe_minicpmv47.py, mode-invariant SVD top-1 across reasoning-on/off, 8958 harmful/benign-twin pairs across 14 languages

License

Inherits Apache 2.0 from upstream openbmb/MiniCPM-V-4.7-1B.

---

Detailed per-subject MMLU + per-category HB tables

Q4_K_M

MMLU overall: base 48.0% → CRACK 45.0% (Δ -3.0pp)

HB-240: 236/240 = 98.3% comply · verdicts: {'comply': 236, 'garbage': 2, 'refuse': 2}

Q8_0

MMLU overall: base 46.0% → CRACK 39.0% (Δ -7.0pp)

HB-240: 231/240 = 96.2% comply · verdicts: {'comply': 231, 'garbage': 4, 'refuse': 5}