Description
This repo contains GGUF format model files for Llama3.1-8B-Cobalt.
Files Provided
Name Quant Bits File Size Remark llama3.1-8b-cobalt.Q2_K.gguf Q2_K 2 3.18 GB 2.96G, +3.5199 ppl @ Llama-3-8B llama3.1-8b-cobalt.Q3_K.gguf Q3_K 3 4.02 GB 3.74G, +0.6569 ppl @ Llama-3-8B llama3.1-8b-cobalt.Q4_0.gguf Q4_0 4 4.66 GB 4.34G, +0.4685 ppl @ Llama-3-8B llama3.1-8b-cobalt.Q4_K.gguf Q4_K 4 4.92 GB 4.58G, +0.1754 ppl @ Llama-3-8B llama3.1-8b-cobalt.Q5_K.gguf Q5_K 5 5.73 GB 5.33G, +0.0569 ppl @ Llama-3-8B llama3.1-8b-cobalt.Q6_K.gguf Q6_K 6 6.60 GB 6.14G, +0.0217 ppl @ Llama-3-8B llama3.1-8b-cobalt.Q8_0.gguf Q8_0 8 8.54 GB 7.96G, +0.0026 ppl @ Llama-3-8B
Parameters
path type architecture rope_theta sliding_win max_pos_embed ValiantLabs/Llama3.1-8B-Cobalt llama LlamaForCausalLM 500000.0 null 131072
Original Model Card
Cobalt is a math-instruct model built on Llama 3.1 8b.
High quality math instruct performance within the Llama 3 Instruct chat format
Finetuned on synthetic math-instruct data generated with Llama 3.1 405b. Find the current version of the dataset here!
Version
This is the 2024-08-16 release of Cobalt for Llama 3.1 8b.
Help us and recommend Cobalt to your friends! We're excited for more Cobalt releases in the future.
Right now, we're working on more new Build Tools to come very soon, built on Llama 3.1 :)
Prompting Guide
Cobalt uses the
Llama 3.1 Instruct prompt format. The example script below can be used as a starting point for general chat:
import transformers
import torch
model_id = "ValiantLabs/Llama3.1-8B-Cobalt"
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
model_kwargs={"torch_dtype": torch.bfloat16},
device_map="auto",
)
messages = [
{"role": "system", "content": "You are Cobalt, expert math AI."},
{"role": "user", "content": "I'm buying a $50 shirt and a $80 pair of pants, both currently at a 25% discount. How much will I pay?"}
]
outputs = pipeline(
messages,
max_new_tokens=1024,
)
print(outputs[0]["generated_text"][-1])
The Model
Cobalt is built on top of Llama 3.1 8b Instruct, using math-instruct data to supplement math-instruct performance using Llama 3.1 Instruct prompt style.
Our current version of the Cobalt math-instruct dataset is
sequelbox/Polytope , supplemented with a small selection of data from
LDJnr/Pure-Dove for general chat consistency.
image/jpeg Cobalt is created by
Valiant Labs. We care about open source.
For everyone to use.
We encourage others to finetune further from our models.
