|
--- |
|
license: creativeml-openrail-m |
|
--- |
|
|
|
Javalion-R is a penta merge of KoboldAI's GPT-J classics + PygmalionAI's Pygmalion6b; |
|
|
|
((Janeway + Shinen) + (Skein + Pygmalion)) + GPT-R. |
|
|
|
Janeway + Shinen is listed under JANIN-GPTJ. Skein + Pygmalion is listed under SKEGMA-GPTJ. |
|
|
|
GPT-R itself is a 60/40 merge of two instruct research models (see digitous/GPT-R for full credits). |
|
|
|
This 5x+ merge is not intended for minors, as it can produce NC-17+ content. |
|
|
|
This model differs from Javelin-R by substituting the Adventure model with Pygmalion, as Adventure is rendered redundant in training data by Skein. |
|
Javalion-R is a research artefact with dual purpose for entertainment as well as an intended example of potential value instruct can bring when combined with models of a different purpose through the use of weight sum merge technology. |
|
|
|
Mileage mat vary. No refunds best wishes. Mainly intended to be utilized with Open Source KoboldAI software. Optimal sampler and settings not determined. Feedback Welcome! |
|
|
|
https://github.com/KoboldAI/KoboldAI-Client |
|
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) |
|
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_digitous__Javalion-R) |
|
|
|
| Metric | Value | |
|
|-----------------------|---------------------------| |
|
| Avg. | 35.42 | |
|
| ARC (25-shot) | 41.72 | |
|
| HellaSwag (10-shot) | 68.02 | |
|
| MMLU (5-shot) | 30.81 | |
|
| TruthfulQA (0-shot) | 34.44 | |
|
| Winogrande (5-shot) | 65.43 | |
|
| GSM8K (5-shot) | 2.65 | |
|
| DROP (3-shot) | 4.85 | |
|
|