This is my "F***ing New Guy" quants for the merges I've been playing with. "I don't know what I'm doing" is an understatement, but until I find an obi-wan, I'm just fumbling around in the dark. Some of the stuff is kinda useable though. Specifically my Stage2-12B seemed hyperfocused on single-persona roles, not wanting to even add in off-the-cuff dialogue. My Stage1 has some weird lightning in a bottle element to it that I can't quite articulate, but it's worth a nudge if you don't mind a little instability. As far as I'm aware theyre all creative commons 4 copyright, but I'll be honest, I was definitely making some of these at 2 am. The quants are only available in imatrix IQ4_XS as that's what I use for testing the temperment of the model since it is quick and easy to make and use on my 6900xt.
Even with public listings being free - I probably won't exacerbate the storage issue and will purge any HF models that don't meet muster. When I get one that I really like I'll go the whole nine yards on quant variety, until then, these are concept cars. Some hot glue is expected.
The following YAML configuration was used to produce these models:
Stage1:
models:
- model: anthracite-org/magnum-v4-12b
parameters:
density: 0.3
weight: 0.3
- model: TheDrummer/Rocinante-12B-v1.1
parameters:
density: 0.7
weight: 0.5
- model: VongolaChouko/Starcannon-Unleashed-12B-v1.0
parameters:
density: 0.1
weight: 0.2
merge_method: ties
base_model: TheDrummer/Rocinante-12B-v1.1
parameters:
normalize: true
int8_mask: true
dtype: bfloat16
Stage2:
base_model: TheDrummer/Rocinante-12B-v1.1
dtype: bfloat16
merge_method: della_linear
parameters:
epsilon: 0.05
int8_mask: 1.0
lambda: 0.8
slices:
- sources:
- layer_range: [0, 40]
model: anthracite-org/magnum-v4-12b
parameters:
density: 0.7
weight: 0.25
- layer_range: [0, 40]
model: nbeerbower/Lyra4-Gutenberg-12B
parameters:
density: 0.6
weight: 0.2
- layer_range: [0, 40]
model: MarinaraSpaghetti/NemoMix-Unleashed-12B
parameters:
density: 0.4
weight: 0.15
- layer_range: [0, 40]
model: TheDrummer/Rocinante-12B-v1.1
parameters:
density: 0.5
weight: 0.4
Stage3:
base_model: TheDrummer/Rocinante-12B-v1.1
dtype: bfloat16
merge_method: della_linear
parameters:
epsilon: 0.05
int8_mask: 1.0
lambda: 0.8
slices:
- sources:
- layer_range: [0, 40]
model: anthracite-org/magnum-v4-12b
parameters:
density: 0.7
weight: 0.20
- layer_range: [0, 40]
model: nbeerbower/Lyra4-Gutenberg-12B
parameters:
density: 0.7
weight: 0.25
- layer_range: [0, 40]
model: MarinaraSpaghetti/NemoMix-Unleashed-12B
parameters:
density: 0.4
weight: 0.15
- layer_range: [0, 40]
model: TheDrummer/Rocinante-12B-v1.1
parameters:
density: 0.5
weight: 0.4
Stage4:
base_model: TheDrummer/Rocinante-12B-v1.1
dtype: bfloat16
merge_method: della_linear
parameters:
epsilon: 0.05
int8_mask: 1.0
lambda: 0.7
slices:
- sources:
- layer_range: [0, 40]
model: anthracite-org/magnum-v4-12b
parameters:
density: 0.3
weight: 0.15
- layer_range: [0, 40]
model: nbeerbower/Lyra4-Gutenberg-12B
parameters:
density: 0.8
weight: 0.3
- layer_range: [0, 40]
model: MarinaraSpaghetti/NemoMix-Unleashed-12B
parameters:
density: 0.3
weight: 0.1
- layer_range: [0, 40]
model: nothingiisreal/MN-12B-Starcannon-v3
parameters:
density: 0.5
weight: 0.1
- layer_range: [0, 40]
model: TheDrummer/Rocinante-12B-v1.1
parameters:
density: 0.5
weight: 0.35
Stage5:
models:
- model: anthracite-org/magnum-v4-12b
parameters:
density: 0.6
weight: 0.3
- model: nbeerbower/Lyra4-Gutenberg-12B
parameters:
density: 0.83
weight: 0.4
merge_method: dare_ties
base_model: TheDrummer/Rocinante-12B-v1.1
parameters:
normalize: true
int8_mask: true
dtype: bfloat16
- Downloads last month
- 75