This repo contains a merge of pre-trained language models created using
mergekit.
A coherent Llama 3 model (composed of fine-tunes based on Instruct) was merged at low weight into a Llama 3.1 Instruct model. No fine-tuning was performed afterward. The resulting model is mostly coherent for direct chat and text generation, retaining long context capability of 3.1.
A gradient merge was used at the ends and the the embed_tokens and lm_head layers retained from 3.1, which should better preserve handling of context above 8K tokens.
Testing has been performed out to 16K context, using temperature 1 and minP 0.01. Safety remains mostly intact.
Built with Llama.
This model was merged using the SLERP merge method.
1base_model: meta-llama/Llama-3.1-8B-Instruct
2dtype: bfloat16
3merge_method: slerp
4slices:
5- sources:
6 - model: meta-llama/Llama-3.1-8B-Instruct
7 layer_range: [0, 32]
8 - model: grimjim/llama-3-Nephilim-v3-8B
9 layer_range: [0, 32]
10 value: [0.0, 0.02, 0.04, 0.06, 0.08, 0.1, 0.1, 0.1,
11 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1,
12 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1,
13 0.1, 0.1, 0.1, 0.08, 0.06, 0.04, 0.02, 0.0]
14parameters:
15 t:
16 - filter: embed_tokens
17 value: 0.0
18 - filter: lm_head
19 value: 0.0
20 - value: 0.1
21
