language:
- en
license: apache-2.0
library_name: transformers
tags:
- mergekit
- merge
base_model:
- sometimesanotion/Qwen2.5-14B-Vimarckoso-v3
- sometimesanotion/Lamarck-14B-v0.3
- sometimesanotion/Qwenvergence-14B-v3-Prose
- Krystalan/DRT-o1-14B
- underwoods/medius-erebus-magnum-14b
- sometimesanotion/Abliterate-Qwenvergence
- huihui-ai/Qwen2.5-14B-Instruct-abliterated-v2
metrics:
- accuracy
pipeline_tag: text-generation
Update: Lamarck has, for the moment, taken the #1 average score for 14 billion parameter models. Counting all the way up to 32 billion parameters, it's #7. This validates the complex merge techniques which captured the complementary strengths of other work in this community. Humor me, I'm giving our guy his meme shades!
Lamarck 14B v0.6: A generalist merge focused on multi-step reasoning, prose, multi-language ability, and code. It is based on components that have punched above their weight in the 14 billion parameter class. Here you can see a comparison between Lamarck and other top-performing merges and finetunes:
Previous releases were based on a SLERP merge of model_stock+della branches focused on reasoning and prose. The prose branch got surprisingly good at reasoning, and the reasoning branch became a strong generalist in its own right. Some of you have already downloaded it as sometimesanotion/Qwen2.5-14B-Vimarckoso-v3.
A notable contribution from the middle to upper layers of Lamarck v0.6 comes from Krystalan/DRT-o1-14B. It has a fascinating research paper: DRT-o1: Optimized Deep Reasoning Translation via Long Chain-of-Thought.
Lamarck 0.6 hit a whole new level of toolchain-automated complexity with its multi-pronged merge strategies:
- Extracted LoRA adapters from special-purpose merges
- Separate branches for breadcrumbs and DELLA merges
- Highly targeted weight/density gradients for every 2-4 layers
- Finalization through SLERP merges recombining the separate branches
This approach selectively merges the strongest aspects of its ancestors. Lamarck v0.6 is my most complex merge to date. The LORA extractions alone pushed my hardware enough to be the building's sole source of heat for several winter days! By comparison, the SLERP merge below which finalized it was a simple step.
name: Lamarck-14B-v0.6-rc4
merge_method: slerp
base_model: sometimesanotion/lamarck-14b-converge-della-linear
tokenizer_source: base
dtype: float32
out_dtype: bfloat16
parameters:
int8_mask: true
normalize: true
rescale: false
parameters:
t:
- value: 0.30
slices:
- sources:
- model: sometimesanotion/lamarck-14b-converge-della-linear
layer_range: [ 0, 8 ]
- model: sometimesanotion/lamarck-14b-converge-breadcrumbs
layer_range: [ 0, 8 ]
- sources:
- model: sometimesanotion/lamarck-14b-converge-della-linear
layer_range: [ 8, 16 ]
- model: sometimesanotion/lamarck-14b-converge-breadcrumbs
layer_range: [ 8, 16 ]
- sources:
- model: sometimesanotion/lamarck-14b-converge-della-linear
layer_range: [ 16, 24 ]
- model: sometimesanotion/lamarck-14b-converge-breadcrumbs
layer_range: [ 16, 24 ]
- sources:
- model: sometimesanotion/lamarck-14b-converge-della-linear
layer_range: [ 24, 32 ]
- model: sometimesanotion/lamarck-14b-converge-breadcrumbs
layer_range: [ 24, 32 ]
- sources:
- model: sometimesanotion/lamarck-14b-converge-della-linear
layer_range: [ 32, 40 ]
- model: sometimesanotion/lamarck-14b-converge-breadcrumbs
layer_range: [ 32, 40 ]
- sources:
- model: sometimesanotion/lamarck-14b-converge-della-linear
layer_range: [ 40, 48 ]
- model: sometimesanotion/lamarck-14b-converge-breadcrumbs
layer_range: [ 40, 48 ]
The strengths Lamarck has combined from its immediate ancestors are in turn derived from select finetunes and merges. Kudoes to @arcee-ai, @CultriX, @sthenno-com, @Krystalan, @underwoods, @VAGOSolutions, and @rombodawg whose models had the most influence. Of this model's immediate ancestors, Vimarckoso v3 has the model card which documents the other finetunes in its extended lineage.