The methodology for training uses PoSE and dynamic-NTK interpolation.
NTK-scaling
The scale factor for NTK is 4. Note that we also tried theta-scaling but this did not work as well as NTK scaling in our experiments.
PoSE
We utilise Positional Skip-wise Training (PoSE) with the following parameters:
Number of Chunks: 5
Max position ID: 32768
Data
We use on average ~8K long samples from RedPajama.
Hardware
We train on 8xH100 GPUs with Deepspeed Zero Stage 3.
Evaluation Methodology
We use the EasyContext implementation of Needle-in-a-Haystack to evaluate Llama-3-Giraffe-70B.
We evaluate with the following parameters:
Min context length: 2000
Max context length: 128000
Context interval: 4000
Depth interval: 0.1
Num samples: 2
Rnd number digits: 7
Haystack dir: PaulGrahamEssays
Adapter Transfer
We apply the above techniques first to Llama-3-70B-Base, using LoRA on the Q and K weights only. This adapter is then applied to Llama-3-70B-Instruct, and we
release the merged version here.
