For this merged model, rope theta was in config.json was manually adjusted down to 100K, a value less than 1M as initially released by Mistral for v0.2, but higher than the 10K that accompanied practical 8K context for v0.1. We idly conjecture that 1M rope theta might improve performance for needle-in-a-haystack queries; however, during informal testing, narrative coherence seemed to occasionally suffer under 1M rope theta. Furthermore, the results reported in the arXiv paper
Scaling Laws of RoPE-based Extrapolation suggest that 1M rope theta may be overkill for a 32K token context window.
Lightly tested with temperature 0.9-1.0 and minP 0.02, using ChatML prompts. The model natively supports Alpaca prompts.
This is a merge of pre-trained language models created using
mergekit.
This model was merged using the SLERP merge method.
1slices:
2 - sources:
3 - model: grimjim/Mistral-Starling-merge-trial1-7B
4 layer_range: [0, 32]
5 - model: grimjim/kukulemon-7B
6 layer_range: [0, 32]
7# or, the equivalent models: syntax:
8# models:
9merge_method: slerp
10base_model: grimjim/Mistral-Starling-merge-trial1-7B
11parameters:
12 t:
13 - filter: self_attn
14 value: [0, 0.5, 0.3, 0.7, 1]
15 - filter: mlp
16 value: [1, 0.5, 0.7, 0.3, 0]
17 - value: 0.5 # fallback for rest of tensors
18dtype: bfloat16
19
