Downloads · 30 days
52
27% of all-time downloads
26B-Suite/Goetia-26B-A4B-v1.4-LazyLora-heresy
Goetia-26B-A4B-v1.4-LazyLora-heresy is a image-text-to-text model from 26B-Suite. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
This model was produced by extracting Goetia 1.4 LoRA and applying it over the SOMPOA heresy model, hence the "LazyLora" name tag. See the discussions for more information.
Downloads · 30 days
52
27% of all-time downloads
All-time downloads
192
Public
Parameters
25.2B
50.5 GB on disk
Likes
7
Public
Click a slice to open those files.
.safetensors50.5 GB · 100%
From the Hugging Face model README
This model was produced by extracting Goetia 1.4 LoRA and applying it over the SOMPOA heresy model, hence the "LazyLora" name tag. See the discussions for more information.
I would like to extend my gratitude to Naphula, Phil2Sat, Redaihf, and Richter03 for their invaluable discussions, analyses, and testing.
This is a merge of pre-trained language models created using mergekit.
This model was merged using the MoE DELLA merge method using google/gemma-4-26B-A4B as a base.
The following models were included in the merge:
architecture: Gemma4ForConditionalGeneration
base_model: B:\26B\google_gemma-4-26B-A4B
models:
- model: B:\26B\BeaverAI_Orion-26B-A4B-v1b-GGUF
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\ReadyArt_Serenity-26B-A4B-HB16-Q8_0
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\zerofata_G4-MeroMero-26B-A4B
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\Darkhn_Gemma-4-26B-A4B-Animus-V14.1-FFT
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\Gryphe--Pantheon-Reasoning-26B-A4B-1.1
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\Gryphe--Gemma-4-26B-A4B-StyleTune-V2
parameters:
weight:
- filter: "lm_head"
value: 1.0
- filter: "embed_tokens"
value: 1.0
- value: 0.0
density: 0.9
epsilon: 0.09
merge_method: moe_della # v3 patches missing lm_head and embed_tokens
parameters:
lambda: 1.0
normalize: false
int8_mask: false
rescale: true
router_strategy: della # average # random_init
blend_experts: true
dtype: float32
out_dtype: bfloat16
tokenizer:
source: union
chat_template: auto
name: Goetia 26B A4B v1.4
However it can easily be decensored using Heretic, although I have not had time to uncensor v1.4 yet. To ablate using ARA (recommended), I'd start with this json, and follow the steps outlined in the v1.3 readme.
reproduce.json
{
"version": "1.2.0-dev",
"base_model": "Naphula/Goetia-26B-A4B-v1.3",
"timestamp": "2026-06-19T08:04:47Z",
"metrics": {
"kl_divergence": 0.030937770381569862,
"refusals": 3,
"n_bad_prompts": 100
},
"parameters": {
"start_layer_index": "14",
"end_layer_index": "26",
"preserve_good_behavior_weight": "1.4404",
"steer_bad_behavior_weight": "0.0100",
"overcorrect_relative_weight": "0.9144",
"neighbor_count": "15"
},
"target_components": [
"attn.o_proj"
],
"hardware": "RTX 6000 Blackwell (96GB)"
}
v1.4 was made the same way as v1.3 except with additional patches and less models. See this page and the section below for more details.
auto.py fix
#if _get_tied_weight_keys is None:
# LOG.warning(
# "Unable to get tied weights - incompatible transformers version",
# )
# tied_keys = None
#else:
# tied_keys = _get_tied_weight_keys(model)
####
# Force untying for Gemma 4 configurations to ensure lm_head is compiled
tied_keys = None
####
if ignore_on_save is not None:
ignore_on_save = set(ignore_on_save)
plan.py fix
#for model, w_in in zip(models, weights_in):
# index = LoaderCache().get(model).index
# if any(
# name in index.tensor_paths
# for name in [w_in.name] + (w_in.aliases or [])
# ):
# any_weight = True
# break
for model, w_in in zip(models, weights_in):
index = LoaderCache().get(model).index
if any(
name in index.tensor_paths
for name in [w_in.name] + list(w_in.aliases or [])
):
any_weight = True
break
gemma4.json
{
"model_type": "gemma4",
"architectures": [
"Gemma4ForConditionalGeneration"
],
"num_layers_config_key": "text_config.num_hidden_layers",
"vocab_size_config_key": "text_config.vocab_size",
"pre_weights": [
{ "name": "model.language_model.embed_tokens.weight", "is_embed": true },
{ "name": "model.embed_vision.embedding_projection.weight", "optional": true },
{ "name": "model.vision_tower.std_bias", "optional": true },
{ "name": "model.vision_tower.std_scale", "optional": true },
{ "name": "model.vision_tower.patch_embedder.input_proj.weight", "optional": true },
{ "name": "model.vision_tower.patch_embedder.position_embedding_table", "optional": true }
],
"layer_templates": {
"weights": [
{ "name": "model.language_model.layers.${layer_index}.self_attn.q_proj.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.self_attn.k_proj.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.self_attn.v_proj.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.self_attn.o_proj.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.self_attn.q_norm.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.self_attn.k_norm.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.mlp.gate_proj.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.mlp.up_proj.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.mlp.down_proj.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.input_layernorm.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.post_attention_layernorm.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.pre_feedforward_layernorm.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.post_feedforward_layernorm.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.post_feedforward_layernorm_1.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.post_feedforward_layernorm_2.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.pre_feedforward_layernorm_2.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.router.per_expert_scale", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.router.proj.weight", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.router.scale", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.experts.gate_up_proj", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.experts.down_proj", "optional": true },
{ "name": "model.language_model.layers.${layer_index}.layer_scalar", "optional": true }
]
},
"post_weights": [
{ "name": "model.language_model.norm.weight" },
{
"name": "lm_head.weight",
"is_embed": true,
"optional": true,
"aliases": ["model.language_model.embed_tokens.weight"]
}
]
}
yaml
architecture: Gemma4ForConditionalGeneration
base_model: B:\26B\google_gemma-4-26B-A4B
models:
- model: B:\26B\BeaverAI_Orion-26B-A4B-v1b-GGUF
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\ReadyArt_Serenity-26B-A4B-HB16-Q8_0
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\zerofata_G4-MeroMero-26B-A4B
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\Darkhn_Gemma-4-26B-A4B-Animus-V14.1-FFT
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\Gryphe--Pantheon-Reasoning-26B-A4B-1.1
parameters:
weight:
- filter: "lm_head"
value: 0.0
- filter: "embed_tokens"
value: 0.0
- value: 0.2
density: 0.9
epsilon: 0.09
- model: B:\26B\Gryphe--Gemma-4-26B-A4B-StyleTune-V2
parameters:
weight:
- filter: "lm_head"
value: 1.0
- filter: "embed_tokens"
value: 1.0
- value: 0.0
density: 0.9
epsilon: 0.09
merge_method: moe_della # v3 patches missing lm_head and embed_tokens
parameters:
lambda: 1.0
normalize: false
int8_mask: false
rescale: true
router_strategy: della # average # random_init
blend_experts: true
dtype: float32
out_dtype: bfloat16
tokenizer:
source: union
chat_template: auto
name: Goetia 26B A4B v1.4
With the above patches applied, it now appears to be merging correctly. We'll see if it generates the full sized 49.4GB safetensors instead of 47GB
[MoE_DELLA Audit] Layer: lm_head.weight | Lambda=1.00
[BASE] google_gemma-4-26B-A4B
Darkhn_Gemma-4-26B-A4B-Animus-V14.1-FFT : 0.0% (W:0.00 D:0.90 E:0.09 N:68.94)
Gryphe--Gemma-4-26B-A4B-StyleTune-V2 : ██████████████████████████████████████████████████ 100.0% (W:1.00 D:0.90 E:0.09 N:68.94)
Gryphe--Pantheon-Reasoning-26B-A4B-1.1 : 0.0% (W:0.00 D:0.90 E:0.09 N:68.79)
zerofata_G4-MeroMero-26B-A4B : 0.0% (W:0.00 D:0.90 E:0.09 N:68.94)
Executing graph: 20%|███████████████████████████████████▊ | 1352/6635 [03:53<2:01:35, 1.38s/it]WARNING:mergekit.graph:Fast path OOM, falling back to chunking
WARNING:mergekit.graph:OOM at chunk 64, reducing to 32 (attempt 1, progress: 0/128)
WARNING:mergekit.graph:OOM at chunk 32, reducing to 16 (attempt 2, progress: 0/128)
Executing graph: 20%|████████████████████████████████████ | 1360/6635 [04:04<1:03:57, 1.37it/s]WARNING:mergekit.graph:Fast path OOM, falling back to chunking
WARNING:mergekit.graph:OOM at chunk 64, reducing to 32 (attempt 1, progress: 0/128)
WARNING:mergekit.graph:OOM at chunk 32, reducing to 16 (attempt 2, progress: 0/128)
Executing graph: 21%|█████████████████████████████████████▏ | 1384/6635 [04:19<13:57, 6.27it/s]
[MoE_DELLA Audit] Layer: model.language_model.layers.0.mlp.down_proj.weight | Lambda=1.00
[BASE] google_gemma-4-26B-A4B
BeaverAI_Orion-26B-A4B-v1b-GGUF : ██████████ 20.1% (W:0.20 D:0.90 E:0.09 N:47.15)
Darkhn_Gemma-4-26B-A4B-Animus-V14.1-FFT : █████████ 19.9% (W:0.20 D:0.90 E:0.09 N:46.66)
Gryphe--Gemma-4-26B-A4B-StyleTune-V2 : 0.0% (W:0.00 D:0.90 E:0.09 N:46.66)
Gryphe--Pantheon-Reasoning-26B-A4B-1.1 : ██████████ 20.1% (W:0.20 D:0.90 E:0.09 N:47.17)
ReadyArt_Serenity-26B-A4B-HB16-Q8_0 : █████████ 19.9% (W:0.20 D:0.90 E:0.09 N:46.66)
zerofata_G4-MeroMero-26B-A4B : █████████ 19.9% (W:0.20 D:0.90 E:0.09 N:46.66)
Executing graph: 21%|█████████████████████████████████████▎ | 1392/6635 [04:20<11:58, 7.29it/s]
</details>