File size: 6,427 Bytes
161638d 1f73368 161638d 1f73368 52e89cd 1f73368 d7c0e06 161638d a75f5b4 15c2263 fd8c232 6b8b883 bd90498 cc4a125 6b8b883 0048610 977f263 1dfe0df 161638d cc4a125 04277a3 9c7cb33 161638d c5c097e 1f73368 161638d c5c097e 161638d cc4a125 977f263 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 |
---
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. With a cut-off of 33 billion parameters, as of this writing, it's currently #7. This validates the complex merge techniques which combined the complementary strengths of other work in this community into one model.
Lamarck 14B v0.6: A generalist merge focused on multi-step reasoning, prose, and multi-language ability. 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](https://huggingface.co/sometimesanotion/Qwen2.5-14B-Vimarckoso-v3).
A notable contribution to the middle to upper layers of Lamarck v0.6 comes from [Krystalan/DRT-o1-14B](https://huggingface.co/Krystalan/DRT-o1-14B). It has a fascinating research paper: [DRT-o1: Optimized Deep Reasoning Translation via Long Chain-of-Thought](https://huggingface.co/papers/2412.17498).
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.
```yaml
---
name: lamarck-14b-v0.6-005-model_stock
merge_method: model_stock
base_model: sometimesanotion/Qwenvergence-14B-Base-v2
tokenizer_source: sometimesanotion/Abliterate-Qwenvergence
dtype: float32
out_dtype: bfloat16
parameters:
int8_mask: true
normalize: true
rescale: false
models:
- model: arcee-ai/Virtuoso-Small-qv64
- model: Krystalan/DRT-o1-14B-qv128
- model: sometimesanotion/Qwen2.5-14B-Vimarckoso-v3-qv64
- model: sometimesanotion/Qwenvergence-14B-v3-Prose-qv256
- model: sometimesanotion/Abliterate-Qwenvergence
---
name: lamarck-14b-converge-breadcrumbs
merge_method: breadcrumbs
base_model: sometimesanotion/lamarck-14b-v0.6-005-model_stock
tokenizer_source: base
dtype: bfloat16
out_dtype: bfloat16
parameters:
int8_mask: true
normalize: true
rescale: false
density: 0.95
weight: 1.00
gamma: 0.018
# Here there be dragons!
---
name: lamarck-14b-converge-della-linear
merge_method: della_linear
base_model: sometimesanotion/Qwen2.5-14B-Vimarckoso-v3
tokenizer_source: base
dtype: float32
out_dtype: bfloat16
parameters:
int8_mask: true
normalize: true
rescale: false
density: 0.95
weight: 1.00
epsilon: 0.018
lambda: 1.20
smoothing_factor: 0.07
# Yep, dragons.
---
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
# Not so dragon-ish.
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](https://huggingface.co/sometimesanotion/Qwen2.5-14B-Vimarckoso-v3) has the model card which documents the other finetunes in its extended lineage. |