llama-3-neural-chat-v2.2-8b

Model Details

Model Description

I fine-tuned llama-3 8B on an approach similar to Intel's neural chat language model. I have slightly modified the data sources so it is stronger in coding, math, and writing. I use both SFT and DPO-Positive. DPO-Positive dramatically improves performance over DPO.

• Developed by: Locutusque
• Model type: Built with Meta Llama 3
• Language(s) (NLP): Many?
• License: Llama 3 license https://huggingface.co/meta-llama/Meta-Llama-3-8B/blob/main/LICENSE

Quants

GGUF: https://huggingface.co/bartowski/llama-3-neural-chat-v2.2-8B-GGUF

ExLlamaV2: https://huggingface.co/bartowski/llama-3-neural-chat-v2.2-8B-exl2

Uses

This model has great performance in writing, coding, and math.

Training Data

Recipe information will be coming soon. This language model's recipe is similar to Intel's Neural Chat.

Direct Use

Conversational AI. This model is also very uncensored, it will respond to pretty much any request regardless of the system prompt, use at your own risk.

Evaluations

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
truthfulqa_mc2	2	none	0	acc	0.5232	±	0.0151
gsm8k	3	strict-match	5	exact_match	0.5974	±	0.0135
		flexible-extract	5	exact_match	0.5974	±	0.0135
agieval_nous	N/A	none	0	acc_norm	0.3841	±	0.0094
		none	0	acc	0.3802	±	0.0094
- agieval_aqua_rat	1	none	0	acc	0.2598	±	0.0276
		none	0	acc_norm	0.2520	±	0.0273
- agieval_logiqa_en	1	none	0	acc	0.3441	±	0.0186
		none	0	acc_norm	0.3687	±	0.0189
- agieval_lsat_ar	1	none	0	acc	0.2217	±	0.0275
		none	0	acc_norm	0.2348	±	0.0280
- agieval_lsat_lr	1	none	0	acc	0.3882	±	0.0216
		none	0	acc_norm	0.3824	±	0.0215
- agieval_lsat_rc	1	none	0	acc	0.4944	±	0.0305
		none	0	acc_norm	0.5019	±	0.0305
- agieval_sat_en	1	none	0	acc	0.6650	±	0.0330
		none	0	acc_norm	0.6553	±	0.0332
- agieval_sat_en_without_passage	1	none	0	acc	0.3981	±	0.0342
		none	0	acc_norm	0.3981	±	0.0342
- agieval_sat_math	1	none	0	acc	0.3500	±	0.0322
		none	0	acc_norm	0.3318	±	0.0318