README.md

---
license: apache-2.0
tags:
- text-generation
- code
- reasoning
- codegemma
- gemma
- safe-tensors
- distillation
- synthetic-dataset
base_model: google/codegemma-1.1-2b
datasets:
- WithinUsAI/GeminiPro3.2_max_distill_god_seed_25k
- WithinUsAI/gemini_3.5_flash_distilled_25k
- WithinUsAI/Gemini_3.2_Pro_Distilled
- WithinUsAI/codegemma_gemini_pro_32_distilled_25k
- WithinUsAI/DEEPMIND_Alpha_Distilled
pipeline_tag: text-generation
library_name: transformers
language:
- en
---

# Gemini3.5-Code.Reasoner-2b-Distilled

Gemini3.5-Code.Reasoner-2b-Distilled is a highly efficient, reasoning-dense model tailored for advanced coding tasks, algorithmic problem-solving, and logical chain-of-thought workflows. 

By applying a specialized Low-Rank Adaptation (LoRA) layer over **CodeGemma 1.1 2B**, this model infuses frontier-level reasoning mechanics into a compact, 2-billion parameter architecture. It bridges the gap between massive cloud-hosted models and local, edge-compute hardware.

## Model Details

- **Developed by:** WithinUsAI
- **Model Type:** Causal Language Model (Fine-tuned / Knowledge Distilled)
- **Base Model:** [google/codegemma-1.1-2b](https://huggingface.co/google/codegemma-1.1-2b)
- **Architecture:** GemmaForCausalLM (CodeGemma variant) + LoRA Adapters
- **License:** Apache 2.0

## Training & Dataset Recipe

The "Reasoner" capabilities of this model are distilled from a multi-source synthetic pipeline focusing on complex coding logic, algorithmic optimization, and step-by-step thinking patterns. The training mixture leverages approximately 100K+ high-quality reasoning examples across five core datasets:

| Dataset Name | Source / Focus | Approx. Size |
| :--- | :--- | :--- |
| `WithinUsAI/GeminiPro3.2_max_distill_god_seed_25k` | High-quality frontier seed prompts for code generation. | ~25k samples |
| `WithinUsAI/gemini_3.5_flash_distilled_25k` | Fast, iterative logical steps and multi-turn debugging data. | ~25k samples |
| `WithinUsAI/Gemini_3.2_Pro_Distilled` | Heavy math logic, structural coding, and system design patterns. | Premium corpus |
| `WithinUsAI/codegemma_gemini_pro_32_distilled_25k` | Target-aligned distillation data optimized for the CodeGemma vocabulary. | ~25k samples |
| `WithinUsAI/DEEPMIND_Alpha_Distilled` | Deep algorithmic competitive programming and math reasoning. | Premium corpus |

## Intended Use

- **Local Code Assistants:** Ideal for IDE plugins requiring fast, low-latency code completion and instruction following.
- **Logical Chain-of-Thought:** Designed to output its reasoning process before writing the final code block, minimizing syntax and logical errors.
- **Resource-Constrained Environments:** Can easily be deployed on mobile devices, single-GPU setups, or local laptops using frameworks like `vLLM`, `Ollama`, or `SGLang`.

## Quickstart Guide

### Inference with Hugging Face Transformers

Because CodeGemma utilizes specialized tokens for coding workflows, it's recommended to structure your prompts cleanly to prompt the model's inner chain-of-thought.

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "WithinUsAI/Gemini3.5-Code.Reasoner-2b-Distilled"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Prompt the model to think step-by-step before delivering code
prompt = """<bos>Analyze the problem and think step-by-step before writing any code.
Problem: Write a Python generator function that yields the Fibonacci sequence up to n elements.
Answer:"""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)

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