NextStep-Coder-MoE / README.md
OsamaBinLikhon's picture
Upload NextStep-Coder-MoE LoRA adapter
23d9c35 verified
metadata
license: apache-2.0
language:
  - en
library_name: peft
tags:
  - base_model:adapter:distilgpt2
  - lora
  - transformers
base_model: distilgpt2
pipeline_tag: text-generation

NextStep-Coder-MoE

A high-performance tiny coding LLM with Interleaved Thinking capability for advanced reasoning and agentic workflows.

Model Description

NextStep-Coder-MoE is a LoRA fine-tuned model based on Qwen2.5-Coder-1.5B, optimized for:

  • Chain-of-Thought reasoning with <think>...</think> tags
  • Multi-step coding tasks
  • Agentic workflows (plan → act → reflect)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load model
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-1.5B")
model = PeftModel.from_pretrained(base_model, "OsamaBinLikhon/NextStep-Coder-MoE")
tokenizer = AutoTokenizer.from_pretrained("OsamaBinLikhon/NextStep-Coder-MoE")

# Generate
prompt = "Write a Python function to check if a number is prime.\n<think>"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0]))

Training Details

Parameter Value
Base Model Qwen/Qwen2.5-Coder-1.5B
Method LoRA (r=16, alpha=32)
Trainable Params 18.4M (1.18%)
Precision bf16
Framework Transformers + PEFT

Interleaved Thinking Format

The model uses <think> tags to show reasoning:

User: Write a binary search function.

Model: <think>
I need to implement binary search on a sorted array.
Key steps: find middle, compare, narrow search space.
Edge case: empty array returns -1.
</think>

def binary_search(arr, target):
    left, right = 0, len(arr) - 1
    while left <= right:
        mid = (left + right) // 2
        if arr[mid] == target:
            return mid
        elif arr[mid] < target:
            left = mid + 1
        else:
            right = mid - 1
    return -1

Intended Use

  • Code generation with step-by-step reasoning
  • Debugging and code review
  • Algorithm design and explanation
  • Educational coding assistance

Limitations

  • Best for Python, may vary for other languages
  • Requires <think> tag retention in conversation history
  • 2048 token context limit

Author

OsamaBinLikhon

License

Apache 2.0

Framework versions

  • PEFT 0.18.0