Enigma is a code-instruct model built on Llama 3.2 3b.
- High quality code instruct performance with the Llama 3.2 Instruct chat format
- Finetuned on synthetic code-instruct data generated with Llama 3.1 405b. Find the current version of the dataset here!
- Overall chat performance supplemented with generalist synthetic data.
Version
This is the 2024-09-30 release of Enigma for Llama 3.2 3b, enhancing code-instruct and general chat capabilities.
Enigma is also available for Llama 3.1 8b!
Help us and recommend Enigma to your friends! We're excited for more Enigma releases in the future.
Prompting Guide
Enigma uses the Llama 3.2 Instruct prompt format. The example script below can be used as a starting point for general chat:
import transformers
import torch
model_id = "ValiantLabs/Llama3.2-3B-Enigma"
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
model_kwargs={"torch_dtype": torch.bfloat16},
device_map="auto",
)
messages = [
{"role": "system", "content": "You are Enigma, a highly capable code assistant."},
{"role": "user", "content": "Can you explain virtualization to me?"}
]
outputs = pipeline(
messages,
max_new_tokens=1024,
)
print(outputs[0]["generated_text"][-1])
The Model
Enigma is built on top of Llama 3.2 3b Instruct, using high quality code-instruct data and general chat data in Llama 3.2 Instruct prompt style to supplement overall performance.
Our current version of Enigma is trained on code-instruct data from sequelbox/Tachibana and general chat data from sequelbox/Supernova.
Enigma is created by Valiant Labs.
Check out our HuggingFace page for Shining Valiant 2 and our other Build Tools models for creators!
Follow us on X for updates on our models!
We care about open source. For everyone to use.
We encourage others to finetune further from our models.
- Downloads last month
- 87
Model tree for ValiantLabs/Llama3.2-3B-Enigma
Base model
meta-llama/Llama-3.2-3B-InstructDatasets used to train ValiantLabs/Llama3.2-3B-Enigma
Collection including ValiantLabs/Llama3.2-3B-Enigma
Evaluation results
- acc on Winogrande (5-Shot)self-reported67.960
- normalized accuracy on ARC Challenge (25-Shot)self-reported47.180
- strict accuracy on IFEval (0-Shot)Open LLM Leaderboard47.750
- normalized accuracy on BBH (3-Shot)Open LLM Leaderboard18.810
- exact match on MATH Lvl 5 (4-Shot)Open LLM Leaderboard6.650
- acc_norm on GPQA (0-shot)Open LLM Leaderboard1.450
- acc_norm on MuSR (0-shot)Open LLM Leaderboard4.540
- accuracy on MMLU-PRO (5-shot)test set Open LLM Leaderboard15.410