Commit
•
ae55d57
1
Parent(s):
8475a53
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
- sr
|
| 6 |
+
- hr
|
| 7 |
+
- bs
|
| 8 |
+
datasets:
|
| 9 |
+
- teknium/OpenHermes-2.5
|
| 10 |
+
- WizardLM/WizardLM_evol_instruct_V2_196k
|
| 11 |
+
- draganjovanovich/airoboros-3.0-serbian
|
| 12 |
+
- Open-Orca/SlimOrca
|
| 13 |
+
- draganjovanovich/prodigy-instruct-reason-function-calling
|
| 14 |
+
- iskonai/chatislav-instruct
|
| 15 |
+
- mlabonne/FineTome-100k
|
| 16 |
+
base_model:
|
| 17 |
+
- draganjovanovich/prodigy-sm-base-v0.1
|
| 18 |
+
---
|
| 19 |
+
# Prodigy SM Instruct v0.1
|
| 20 |
+
|
| 21 |
+
<img src="https://cdn-uploads.huggingface.co/production/uploads/617bbeec14572ebe9e6ea83f/5sl1RHE054rNkJMA4Z0dp.png" width="70%" height="70%">
|
| 22 |
+
|
| 23 |
+
**Prodigy SM Instruct v0.1** is an instruction-tuned version of our [Prodigy SM Base v0.1](https://huggingface.co/iskonai/prodigy-sm-base-v0.1) model, optimized for instruction following in Serbian, Croatian, Bosnian and English languages. The model demonstrates exceptional capabilities in function calling and tool usage while maintaining strong performance across all supported languages.
|
| 24 |
+
|
| 25 |
+
Model was shared as part of the presenatation at DSC Europe 2024 / Belgrade (BREAKOUT SESSIONS & WORKSHOPS).
|
| 26 |
+
|
| 27 |
+
# 🎮 DEMO TIME! 🚀 For next couple of weeks you can test model at: [https://chat.iskon.ai](https://chat.iskon.ai)
|
| 28 |
+
or you can download it and use **SGLang**(recomended as it have superior structured output sampling), vLLM, HF text-generation-inference etc.
|
| 29 |
+
proper model card coming soon when we upload final checkpoint, and remove **draft** from the name :)
|
| 30 |
+
|
| 31 |
+
# Training
|
| 32 |
+
|
| 33 |
+
The model was instruction-tuned using a carefully curated mix of high-quality datasets in ChatML format.
|
| 34 |
+
|
| 35 |
+
# Features
|
| 36 |
+
- Precise **system prompt** following capabilities in **Serbian** (trained on really large system prompts: 500+ tokens.)
|
| 37 |
+
- Strong instruction following capabilities in Serbian, Croatian, Bosnian and English
|
| 38 |
+
- Enhanced function calling and tool usage abilities
|
| 39 |
+
- Maintains the base model's strong performance in all supported languages
|
| 40 |
+
- Uses ChatML format for conversations
|
| 41 |
+
|
| 42 |
+
# Use Cases
|
| 43 |
+
|
| 44 |
+
The model is particularly well-suited for:
|
| 45 |
+
- Multilingual applications requiring Serbian, Croatian, Bosnian and English support
|
| 46 |
+
- Applications requiring precise instruction following
|
| 47 |
+
- Systems utilizing function calling and tool usage
|
| 48 |
+
- Conversational AI applications
|
| 49 |
+
- Text generation in supported languages
|
| 50 |
+
|
| 51 |
+
# Limitations
|
| 52 |
+
|
| 53 |
+
As with all language models:
|
| 54 |
+
- Outputs should be verified for critical applications
|
| 55 |
+
- Performance in languages other than Serbian, Croatian, Bosnian and English may vary
|
| 56 |
+
- Users should be aware of general LLM limitations regarding hallucination and factual accuracy
|
| 57 |
+
|
| 58 |
+
# What's Next
|
| 59 |
+
|
| 60 |
+
Stay tuned for our upcoming releases:
|
| 61 |
+
- [prodigy-sm-base (llama-3.1)]() **COMING SOON**
|
| 62 |
+
- [prodigy-sm-instruct (llama-3.1)]() **COMING SOON**
|
| 63 |
+
|
| 64 |
+
📢 We're excited to announce that [iskon.ai](https://Iskon.ai) will soon launch an API platform featuring advanced **Prodigy** series of models, advanced AI tools and much more! 🚀
|