| --- |
| 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)) |
| |
