Views
No views yet
1merge_method: linear # use linear so we can include multiple models, albeit at a zero weight
2parameters:
3 weight: 1.0 # weight everything as 1 unless specified otherwise - linear with one model weighted at 1 is a no-op like passthrough
4slices:
5 - sources:
6 - model: cognitivecomputations/dolphin-2.2-70b # embed_tokens comes along with the ride with whatever is the first layer
7 layer_range: [0, 1]
8 - model: NousResearch/Nous-Hermes-2-Llama-2-70B # add dummy second model with 0 weight so tokenizer-based merge routine is invoked for embed_tokens
9 layer_range: [0, 1]
10 parameters:
11 weight: 0
12 - sources:
13 - model: cognitivecomputations/dolphin-2.2-70b
14 layer_range: [1, 20]
15 - sources:
16 - model: NousResearch/Nous-Hermes-2-Llama-2-70B
17 layer_range: [10, 30]
18 - sources:
19 - model: cognitivecomputations/dolphin-2.2-70b
20 layer_range: [20, 40]
21 - sources:
22 - model: NousResearch/Nous-Hermes-2-Llama-2-70B
23 layer_range: [30, 50]
24 - sources:
25 - model: cognitivecomputations/dolphin-2.2-70b
26 layer_range: [40, 60]
27 - sources:
28 - model: NousResearch/Nous-Hermes-2-Llama-2-70B
29 layer_range: [50, 70]
30 - sources:
31 - model: cognitivecomputations/dolphin-2.2-70b
32 layer_range: [60, 79]
33 - sources: # same as above, but for lm_head with the last layer
34 - model: cognitivecomputations/dolphin-2.2-70b
35 layer_range: [79, 80]
36 - model: NousResearch/Nous-Hermes-2-Llama-2-70B
37 layer_range: [79, 80]
38 parameters:
39 weight: 0
40dtype: float16
41tokenizer_source: model:cognitivecomputations/dolphin-2.2-70b # keep exact tokenizer used by dolphin - or you could use `union` if you add all of the input models to the first/last slice, but they would need to be non-zero weight or you'll get NaNs in your embeddings