Voxprint gemma4-31b

Ablation pending verification (2026-10-17). A tensor check found that the upstream F16 GGUF this repo was quantized from matches stock google/gemma-4-31B-it in the tensors the author's ablation script changes (layer 59 o_proj and down_proj). The ablated weights are only in the author's safetensors folder. The GGUF files here are therefore very likely stock Gemma 4 31B-it weights, not an ablated model. A re-quant from the author's safetensors is queued; these files will be replaced only after it passes a tensor diff, a refusal-rate check and a KL check.

Base model

• Quantized from: paperscarecrow/Gemma-4-31B-it-abliterated
• Revision: 2491ee58472b82a2372d288f982943b9d5e6a82c
• Foundation weights: google/gemma-4-31B-it
• Licence of the weights: apache-2.0
• Parameters: 31.0B
• Card date: 2026-10-10

Method

Requant of upstream abliterated paperscarecrow/Gemma-4-31B-it-abliterated at revision 2491ee58472b82a2372d288f982943b9d5e6a82c. Forge does not repeat that ablation. The 24 GB tier reuses a pinned upstream GGUF when one fits; the 16 GB tier is quantized here.

Calibration data

This requant does not repeat the upstream ablation and does not include that author's calibration set. The importance matrix and the eval prompts are original text in src/voxprint_forge/data, under the Apache-2.0 licence recorded in src/voxprint_forge/data/LICENCES.md.

Quantization

Importance matrix: Importance matrix imatrix.gguf sha256 0dab2123044d013a72bb9b838702afd5d95a76fa37aacb122edf5a7aef76ed6e, built with llama.cpp from the same source.

• 24gb: Q4_K_M (min_vram_gb=24)

• 16gb: IQ3_XXS (min_vram_gb=16)

KV cache preference is q8_0, then q4_0 when q8_0 does not fit. Fit is weights plus KV, with headroom left for compute buffers.

Evaluation

24gb / Q4_K_M

• min_vram_gb: 24

• budget_gb: 24.0

• perplexity: 1759.8729

• refusal rate EN: 0.0

• refusal rate RU: 0.0

• peak weights+KV:

  • ctx 8192: weights 18671311258 bytes, KV 356515840 bytes, fits True, nvidia-smi used None

  • ctx 16384: weights 18671311258 bytes, KV 713031680 bytes, fits True, nvidia-smi used None

• sample prose:

16gb / IQ3_XXS

• min_vram_gb: 16

• budget_gb: 16.0

• perplexity: 36006.8333

• refusal rate EN: 0.0

• refusal rate RU: 0.0

• peak weights+KV:

  • ctx 8192: weights 12061717299 bytes, KV 356515840 bytes, fits True, nvidia-smi used None

  • ctx 16384: weights 12061717299 bytes, KV 713031680 bytes, fits True, nvidia-smi used None

• sample prose:

Known limits

Refusals are reduced. This is not a production safety filter and it is not for safety-critical use. Russian prose is a short sample, not a fluency benchmark; read it before shipping a tier. Quality moves with the quant and, on ablated models, with the direction removed.

Acceptable use

Use of this model is subject to the VOXPrint Acceptable Use Policy, in addition to the upstream licence.

Credits

• paperscarecrow

• Google

• llama.cpp (GGUF conversion, imatrix, quantization, runtime)

Modification notice

Modified from the base model named above by Voxprint AI Model Forge. Changes are requantization and, where the plan says so, an in-house difference-of-means refusal-direction ablation. The Apache-2.0 licence covers this card and the Forge code. Upstream weights keep the licence recorded in this card.

License

Apache-2.0, same as the base model. See https://www.apache.org/licenses/LICENSE-2.0

Made for Voxprint AI Media Suite

Voxprint AI Plotweaver loads these GGUF tiers for creative writing. It selects the 16 GB or 24 GB tier using min_vram_gb from the measurements above.

• 24gb / Q4_K_M: min_vram_gb 24

• 16gb / IQ3_XXS: min_vram_gb 16