Edit model card

Koss-7B

Training Time: 1.85h


Model Average ⬆️ ARC HellaSwag MMLU TruthfulQA
NewstaR/Koss-7B-chat 📑 55.79 53.67 78.79 46.72 43.97

Koss-7B is the smallest variant in the Koss series of neural network models developed by Kaleido AI for natural language processing. With 7 billion parameters, it retains much of the architecture and capabilities of the larger Koss models but requires less computation to run.

Koss-7B is intended for general NLP applications including text classification, language generation, question answering, translation, and dialogue. Its small size makes it suitable for applications with constraints on memory, compute, latency, or carbon emissions.

Factors:

  • Koss-7B should not be used for tasks requiring very specialized knowledge or skills, since its limited parameters reduce expertise in niche domains. For best performance, finetune on in-domain data.
  • As with all AI systems, Koss-7B's behavior is dependent on its training data. It may exhibit biases inherited from non-diverse data. Audit data and mitigation strategies to avoid unfair impacts.
  • Koss-7B is not a creative agent. Its outputs will be limited to recombinations of patterns in its training data. Do not ascribe human-like agency or consciousness.

Recommended Prompt Template:

<s>[INST] {prompt} [/INST] {response} </s>
or
<s>[INST] {prompt} [/INST] 

The model will start it's response after the [/INST] Example:

<s>[INST] Why did the chicken cross the road? [/INST] To get to the other side! </s>

Loss

Loss.png

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 44.98
ARC (25-shot) 53.67
HellaSwag (10-shot) 78.79
MMLU (5-shot) 46.72
TruthfulQA (0-shot) 43.97
Winogrande (5-shot) 71.74
GSM8K (5-shot) 7.35
DROP (3-shot) 12.62
Downloads last month
716
Inference API
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train NewstaR/Koss-7B-chat