naazimsnh02/gemma-4-e4b-opus-reasoning-v2

Overview

This model, naazimsnh02/gemma-4-e4b-opus-reasoning-v2, is a reasoning-enhanced fine-tune of Google's gemma-4-E4B-it base model, which has 4.5 billion effective parameters (8B with embeddings) and a 128K context window. It was fine-tuned using LoRA via Unsloth, specifically targeting attention and MLP layers, and merged in float16 precision.

Key Capabilities & Training

The model's primary focus is on teaching structured reasoning. It was trained on approximately 20,000 samples, combining reasoning distillation from Claude Opus 4.6 traces and supplementary math Chain-of-Thought (CoT) data, with about 40% math content. A key aspect of its training involved formatting assistant responses with <think>...</think> blocks to encourage the model to externalize its reasoning process before providing an answer. This approach aims to teach a reasoning style rather than guaranteeing improved accuracy on specific tasks or benchmarks.

Limitations & Intended Use

It's important to note that this model is reasoning-focused, not benchmark-optimized, and has not been evaluated on standard benchmarks like MMLU or GSM8K. Its reasoning style is derived from Claude Opus 4.6, which may introduce distillation artifacts. As a small model (4.5B effective parameters), it has inherent capacity limits for complex multi-step reasoning or knowledge-intensive tasks. It is not safety-tuned beyond the base model and is predominantly English-only.

Good for:

• Research and experimentation with reasoning distillation techniques.
• Exploring chain-of-thought behavior in smaller language models.
• Personal and educational projects requiring a lightweight reasoning model.
• As a starting point for further fine-tuning efforts.