TinyLlama-1.1bee 🐝

As we feverishly hit the refresh button on hf.co's homepage, on the hunt for the newest waifu chatbot to grace the AI stage, an epiphany struck us like a bee sting. What could we offer to the hive-mind of the community? The answer was as clear as honey—beekeeping, naturally. And thus, this un-bee-lievable model was born.

Details

This model is a fine-tuned version of PY007/TinyLlama-1.1B-intermediate-step-240k-503b on the BEE-spoke-data/bees-internal dataset. It achieves the following results on the evaluation set:

• Loss: 2.4285
• Accuracy: 0.4969

***** eval metrics *****
  eval_accuracy           =     0.4972                                              
  eval_loss               =     2.4283
  eval_runtime            = 0:00:53.12
  eval_samples            =        239
  eval_samples_per_second =      4.499
  eval_steps_per_second   =      1.129
  perplexity              =    11.3391

📜 Intended Uses & Limitations 📜

Intended Uses:

1. Educational Engagement: Whether you're a novice beekeeper, an enthusiast, or someone just looking to understand the buzz around bees, this model aims to serve as an informative and entertaining resource.
2. General Queries: Have questions about hive management, bee species, or honey extraction? Feel free to consult the model for general insights.
3. Academic & Research Inspiration: If you're diving into the world of apiculture studies or environmental science, our model could offer some preliminary insights and ideas.

Limitations:

1. Not a Beekeeping Expert: As much as we admire bees and their hard work, this model is not a certified apiculturist. Please consult professional beekeeping resources or experts for serious decisions related to hive management, bee health, and honey production.
2. Licensing: Apache-2.0, following TinyLlama
3. Infallibility: Our model can err, just like any other piece of technology (or bee). Always double-check the information before applying it to your own hive or research.
4. Ethical Constraints: This model may not be used for any illegal or unethical activities, including but not limited to: bioterrorism & standard terrorism, harassment, or spreading disinformation.

Training and evaluation data

While the full dataset is not yet complete and therefore not yet released for "safety reasons", you can check out a preliminary sample at: bees-v0

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 4
• eval_batch_size: 4
• seed: 80085
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 2.0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	29.15
ARC (25-shot)	30.55
HellaSwag (10-shot)	51.8
MMLU (5-shot)	24.25
TruthfulQA (0-shot)	39.01
Winogrande (5-shot)	54.46
GSM8K (5-shot)	0.23
DROP (3-shot)	3.74