Llama-3-Giraffe-70B-Instruct

Abacus.AI presents our longer-necked variant of Llama 3 70B - now with the instruct variant!

This model has an effective context length of approximately 128k.

We have currently trained on ~1.5B tokens.

There are our Needle-in-a-Haystack heatmap results. We are conducting further evals of model efficacy and will update our model card as these come in:

MT-Bench Evaluation

We also measured performance on MT-Bench to verify that the context extension did not significantly impact performance on instruct tasks:

####### 1st turn:
Meta-Llama-3-70B-Instruct      9.21
Llama-3-Giraffe-70B-Instruct 9.19

####### 2nd turn:
Meta-Llama-3-70B-Instruct     2   8.80
Llama-3-Giraffe-70B-Instruct 2   8.54

####### average:
Meta-Llama-3-70B-Instruct      9.00
Llama-3-Giraffe-70B-Instruct 8.87 

Training Methodology

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.