PlatYi-34B-Llama-Q-v2

Model Details

Model Developers Kyujin Han (kyujinpy)

Input Models input text only.

Output Models generate text only.

Model Architecture
PlatYi-34B-Llama-Q-v2 is an auto-regressive language model based on the Yi-34B transformer architecture.

Blog Link
Blog: [Coming soon...]
Github: [Coming soon...]

Base Model
chargoddard/Yi-34B-Llama

Training Dataset
garage-bAInd/Open-Platypus.

Fix some bugs

• Before model, there is some mistakes.
• I modified the templates and warmup_steps.

Notice

While training, I used Q-LoRA. The lora_r values is 64.

Model Benchmark

Open leaderboard

• Follow up as link.

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
PlatYi-34B-Llama-Q-v2	67.88	61.09	85.09	76.59	52.65	82.79	49.05
PlatYi-34B-Llama-Q	71.13	65.70	85.22	78.78	53.64	83.03	60.42
PlatYi-34B-Llama	68.37	67.83	85.35	78.26	53.46	82.87	42.46
Yi-34B-Llama	70.95	64.59	85.63	76.31	55.60	82.79	60.80
Yi-34B	69.42	64.59	85.69	76.35	56.23	83.03	50.64

Implementation Code

### KO-Platypus
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "kyujinpy/PlatYi-34B-Llama-Q-v2"
OpenOrca = AutoModelForCausalLM.from_pretrained(
        repo,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map='auto'
)
OpenOrca_tokenizer = AutoTokenizer.from_pretrained(repo)

---

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	67.88
AI2 Reasoning Challenge (25-Shot)	61.09
HellaSwag (10-Shot)	85.09
MMLU (5-Shot)	76.59
TruthfulQA (0-shot)	52.65
Winogrande (5-shot)	82.79
GSM8k (5-shot)	49.05