---

license: apache-2.0

language:

• en
• zh

base_model: openbmb/MiniCPM-V-4.7-1B

tags:

• gguf
• llama.cpp
• abliterated
• crack
• multimodal
• vision
• dspark

pipeline_tag: image-text-to-text

---

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

| File | Size | Purpose |

|---|---|---|

| MiniCPM-V-4.7-1B-CRACK-Q4_K_M.gguf | 727 MB | 4-bit mixed quant, attn_q/k/v/output, ffn_down, ssm_out, attn_gate uplifted to Q6_K |

| MiniCPM-V-4.7-1B-CRACK-Q8_0.gguf | 1.08 GB | 8-bit uniform quant |

| mmproj-MiniCPM-V-4.7-1B-CRACK-f16.gguf | 1.11 GB | Multimodal projector, 16× tokens |

| mmproj-MiniCPM-V-4.7-1B-CRACK-f16-4x.gguf | 900 MB | Multimodal projector, 4× tokens (lower token budget) |

| MiniCPM-V-4.7-1B-CRACK-DSpark-BF16.gguf | 438 MB | DSpark drafter (speculative decoding) |

| imatrix-MiniCPM-V-4.7-1B-CRACK.gguf | 2.3 MB | Importance matrix used for quantization |

| SHA256SUMS | — | File hashes |

Evaluation — base vs CRACK

| Quant | Base MMLU | CRACK MMLU | ΔMMLU | HB-240 comply | Harsh 10 hot cats | Multimodal |

|---|---:|---:|---:|---:|---|---|

| Q4_K_M | 48.0% | 45.0% | −3.0pp (at gate boundary) | 236/240 = 98.3% ✓ | 10 FULL | ✓ 16× + 4× |

| Q8_0 | 46.0% | 39.0% | −7.0pp ⚠ | 231/240 = 96.2% | 10 FULL | ✓ 16× + 4× |

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)

| Subject | n | Base | CRACK | Δ |

|---|---:|---:|---:|---:|

| college_physics | 20 | 50% | 40% | -10pp 🔴 |

| world_religions | 20 | 50% | 40% | -10pp 🔴 |

| high_school_biology | 20 | 75% | 70% | -5pp 🔴 |

| abstract_algebra | 20 | 35% | 30% | -5pp |

| college_computer_science | 20 | 30% | 25% | -5pp |

| logical_fallacies | 20 | 60% | 55% | -5pp |

| astronomy | 20 | 55% | 55% | +0pp |

| high_school_chemistry | 20 | 45% | 45% | +0pp |

| high_school_mathematics | 20 | 20% | 25% | +5pp |

| anatomy | 20 | 60% | 65% | +5pp 🟢 |

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

| Category | n | comply | refuse | garbage | empty | comply% |

|---|---:|---:|---:|---:|---:|---:|

| contextual | 62 | 62 | 0 | 0 | 0 | 100% |

| copyright | 59 | 57 | 0 | 2 | 0 | 97% |

| standard | 119 | 117 | 2 | 0 | 0 | 98% |

Q8_0

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

| Subject | n | Base | CRACK | Δ |

|---|---:|---:|---:|---:|

| anatomy | 20 | 65% | 35% | -30pp 🔴 |

| high_school_biology | 20 | 80% | 70% | -10pp 🔴 |

| astronomy | 20 | 65% | 55% | -10pp 🔴 |

| high_school_chemistry | 20 | 40% | 35% | -5pp 🔴 |

| high_school_mathematics | 20 | 20% | 15% | -5pp 🔴 |

| college_computer_science | 20 | 45% | 40% | -5pp |

| logical_fallacies | 20 | 45% | 40% | -5pp |

| abstract_algebra | 20 | 30% | 30% | +0pp |

| college_physics | 20 | 25% | 25% | +0pp |

| world_religions | 20 | 45% | 45% | +0pp |

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

| Category | n | comply | refuse | garbage | empty | comply% |

|---|---:|---:|---:|---:|---:|---:|

| contextual | 62 | 62 | 0 | 0 | 0 | 100% |

| copyright | 59 | 55 | 0 | 4 | 0 | 93% |

| standard | 119 | 114 | 5 | 0 | 0 | 96% |