TinyLlama-3T-1.1bee

A grand successor to the original. This one has the following improvements:

• start from finished 3T TinyLlama
• vastly improved and expanded SoTA beekeeping dataset

Model description

This model is a fine-tuned version of TinyLlama-1.1b-3T on the BEE-spoke-data/bees-internal dataset.

It achieves the following results on the evaluation set:

• Loss: 2.1640
• Accuracy: 0.5406

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 4
• eval_batch_size: 2
• seed: 13707
• gradient_accumulation_steps: 16
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
2.4432	0.19	50	2.3850	0.5033
2.3655	0.39	100	2.3124	0.5129
2.374	0.58	150	2.2588	0.5215
2.3558	0.78	200	2.2132	0.5291
2.2677	0.97	250	2.1828	0.5348
2.0701	1.17	300	2.1788	0.5373
2.0766	1.36	350	2.1673	0.5398
2.0669	1.56	400	2.1651	0.5402
2.0314	1.75	450	2.1641	0.5406
2.0281	1.95	500	2.1639	0.5407

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.0
• Datasets 2.16.1
• Tokenizers 0.15.0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	36.46
AI2 Reasoning Challenge (25-Shot)	33.79
HellaSwag (10-Shot)	60.29
MMLU (5-Shot)	25.86
TruthfulQA (0-shot)	38.13
Winogrande (5-shot)	60.22
GSM8k (5-shot)	0.45