README.md

metadata

license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
tags:
  - mixture-of-experts
  - moe
  - code
  - function-calling
  - agentic
  - qwen2.5
language:
  - en
pipeline_tag: text-generation
library_name: transformers

Fyodor Agentic v1.1

A Mixture-of-Experts (MoE) enhanced version of Qwen2.5-Coder-3B-Instruct, optimized for agentic AI workflows and function calling.

Model Details

• Base Model: Qwen/Qwen2.5-Coder-3B-Instruct
• Architecture: Sparse MoE with 4 experts, top-2 routing
• Parameters: 6-7B total (~3B base + ~3-4B MoE experts)
• Training: Function calling, Python code, conversational data
• Format: SafeTensors (ready to use!)

Quick Start

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "Kiy-K/fyodor-agentic-v1.1",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(
    "Kiy-K/fyodor-agentic-v1.1",
    trust_remote_code=True
)

# Generate
prompt = "Write a Python function to calculate Fibonacci numbers:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Use Cases

• Function Calling: Built for tool usage and API interactions
• Code Generation: Python, JavaScript, and more
• Agentic Workflows: Multi-step reasoning and planning
• Conversational AI: Natural multi-turn dialogue
• Instruction Following: Clear and precise responses

Training Details

• Total Steps: 5000
• Final Loss: 5.5744
• Best Loss: 5.5627
• Training Time: 0.31h
• Platform: Lightning.AI A100

Training Data:

• xLAM Function Calling: 5,000 samples
• Python Code: 5,000 samples
• UltraChat: 5,000 samples

Architecture

Sparse MoE implementation:

• MoE layers added every 3 transformer layers
• 4 experts per layer
• Top-2 routing (2 experts active per token)
• Load balancing for efficient utilization
• Base model frozen, MoE trained

Tips

• Temperature: 0.7-0.8 for creative, 0.3-0.5 for precise
• Context Window: 2048 tokens
• Batch Size: Supports efficient batch inference
• Precision: Works with fp16/bf16/fp32

Advanced Usage

# With custom generation config
outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.8,
    top_p=0.95,
    top_k=50,
    repetition_penalty=1.1,
    do_sample=True
)

License

Apache 2.0 (inherited from base model)

Acknowledgments

• Qwen Team for the excellent base model
• Lightning.AI for compute infrastructure
• HuggingFace for model hosting

---

Built with love for the agentic AI community