Shokey Ai
🧠 Overview
...
Shokey Ai Features
Shokey Ai is an experimental LoRA adapter based on
Qwen2.5-7B-Instruct.
The model is developed by Shokey Corporation and focuses on
reasoning, programming, structured problem solving and natural
conversation.
✨ Features
Reasoning & Analysis
The model is optimized for:
logical analysis;
multi-step problem solving;
decision making;
comparing different solutions;
explaining conclusions.
Programming
Supported tasks:
code generation;
debugging;
architecture planning;
API design;
technical explanations.
Conversation
The model supports:
natural conversations;
roleplay scenarios;
creative writing;
assistant-style interaction.
🏋️ Training
The model was trained using:
LoRA (Low-Rank Adaptation)
with:
Unsloth
4-bit QLoRA training
Hugging Face PEFT
Dataset Composition
Training data consisted of several sources:
reasoning_pairs
Instruction → answer pairs focused on:
analytical tasks;
logical reasoning;
structured answers.
sft_conversations
Multi-turn conversations transformed into supervised fine-tuning examples.
code_corpus
Programming examples with synthetic instructions.
⚙️ Training Configuration
| Parameter | Value |
|---|
| Base model | unsloth/Qwen2.5-7B-Instruct-bnb-4bit |
| Method | LoRA |
| Rank | r=16 |
| Alpha | 32 |
| Dropout | 0 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Trainable parameters | 40,370,176 |
| Percentage | ~0.53% |
| Sequence length | 1024-2048 |
| Optimizer | adamw_8bit |
| Learning rate | 2e-4 |
| Hardware | Tesla T4 (Google Colab Free Tier) |
💻 Usage
Example loading with Unsloth:
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="unsloth/Qwen2.5-7B-Instruct-bnb-4bit",
max_seq_length=2048,
load_in_4bit=True,
)
model.load_adapter(
"Sh0key/Shokey-Ai-LoRA"
)
⚠️ Limitations
This is an early experimental release.
Known issues:
Overly detailed responses
The model may generate very long structured answers even for simple questions.
Example style:
Analysis
↓
Options
↓
Advantages
↓
Disadvantages
↓
Conclusion
Possible hallucination about actions
The model may sometimes claim that it:
modified files;
changed code;
completed actions;
when it only generated an explanation.
Always verify generated results.
No native reasoning mode
This model does not include the native "thinking mode" feature introduced in newer Qwen3-based models.
📜 License
This adapter is released under:
CC-BY-NC-4.0
Allowed:
✅ Personal use
✅ Research
✅ Educational projects
Not allowed:
❌ Commercial usage without permission
The original Qwen2.5 model is distributed separately under:
🏢 Credits
Created by:
Shokey Corporation
AI Research Division
Project:
Shokey Ai
Building experimental AI systems for the future.
