Here is my custom mixed-quant GGUF, which I use regularly. It fits fine into 16GB VRAM with a 16K context window (using Q8 KV cache). If you need mmproj, it's available here.

GGUF recipe

llama-quantize \
  --imatrix imatrix.gguf \
  --token-embedding-type q8_0 \
  --output-tensor-type q8_0 \
  --tensor-type ".*attn_q.weight=q8_0" \
  --tensor-type ".*attn_k.weight=q8_0" \
  --tensor-type ".*attn_output.weight=q5_k" \
  --tensor-type ".*attn_v.weight=iq4_nl" \
  --tensor-type ".*ffn_up.weight=iq4_nl" \
  --tensor-type ".*ffn_gate.weight=iq4_nl" \
  Ministral-3-14B-Nymphaea-RP.F16.gguf \
  Ministral-3-14B-Nymphaea-RP.Q5_Mix.gguf \
  q5_k

Imatrix file for making your own quants is available here. I used this calibration dataset to create it, expanding it with RP and creative writing data (about 400k tokens).

Training Notes

Trained on the latest iteration of my Darkmere dataset. This version features expanded genre variety, built upon a mix of manually curated synthetics and human-written stories.

[!IMPORTANT] The base weights are abliterated via Heretic prior to fine-tuning, so this fine-tune is quite uncensored.

Training Specs

Method:

Training Method: DoRA (Weight-Decomposed LoRA)
Target Modules all-linear
LoRA Rank: 64
LoRA Alpha: 64
LoRA Dropout: 0.05

Hyperparameters:

Batch Size: 2 (Per-device)
Gradient Accumulation: 2
Epochs: 2
Learning Rate: 1e-4
Optimizer: adamw_torch_fused
LR Scheduler: cosine
Noise Level: neftune_noise_alpha=5

[!Note] The vision encoder was frozen during training, so the model retains its native vision capabilities.

Special Thanks

This fine-tune wouldn't be possible without the incredible work of the community:

p-e-w for developing Heretic - an essential tool for censorship removal.
SicariusSicariiStuff for developing SLOP_Detector script.
Mistral AI for their Ministral 3 weights.
AMD for their Instinct™ MI300X GPU.

0xA50C1A1/Ministral-3-14B-Nymphaea-RP

Ministral-3-14B-Nymphaea-RP

GGUF

Training Notes

Special Thanks

0xA50C1A1/
Ministral-3-14B-Nymphaea-RP