Lucid-Research/LucentLogico-3B

LucentLogico

LucentLogico is a family of compact reasoning-specialized language models developed by Lucent Research. Each model is fine-tuned from IBM Granite 4.0 base architectures and optimized for structured, multi-step analytical reasoning.

The LucentLogico series focuses on mathematical derivation, algorithmic code reasoning, and formal logic tasks, with explicit intermediate reasoning steps emphasized during training.

---

Model Variants

LucentLogico-3B

• Base: ibm-granite/granite-4.0-micro-base
• Parameter Class: ~3B
• Target Use: High-capacity compact reasoning systems

LucentLogico-1B

• Base: ibm-granite/granite-4.0-1b-base
• Parameter Class: ~1B
• Target Use: Efficient reasoning with reduced compute requirements

LucentLogico-350M

• Base: ibm-granite/granite-4.0-350m-base
• Parameter Class: ~350M
• Target Use: Lightweight reasoning experimentation and edge deployment

All variants are trained using the same reasoning-focused dataset and training philosophy, scaled to their respective parameter sizes.

---

Training Dataset

All LucentLogico models were fine-tuned on:

Lucid-Research/advanced-reasoning-v1-smol
36,000 curated instruction–response pairs dedicated exclusively to advanced reasoning.

Dataset Composition

The dataset is a balanced tri-domain blend.

Mathematical Reasoning (12,000 samples)

Source: MetaMathQA

• Multi-step mathematical derivations
• Symbolic manipulation
• Competition-style reasoning problems
• Explicit step-by-step solutions

Code & Algorithmic Reasoning (12,000 samples)

Sources: Magicoder-OSS-Instruct-75K, CodeAlpaca-20k

• Natural language specification to code
• Algorithm design tasks
• Debugging and refinement examples
• Structured execution planning

Formal Logic & STEM Reasoning (12,000 samples)

Source: SlimOrca

• Logic puzzles
• Proof-style reasoning
• Scientific inference
• Multi-hop structured deduction

---

Design Principles

The LucentLogico series was trained with the following priorities:

• Explicit intermediate reasoning in every example
• Balanced cross-domain analytical capability
• Reduced reasoning drift
• Structured decomposition of complex problems
• Standardized instruction–response formatting

The training dataset deliberately excludes conversational and alignment-focused data in order to maintain strict specialization in reasoning performance.

---

Intended Use

LucentLogico models are designed for:

• Step-by-step mathematical reasoning
• Algorithmic code synthesis
• Logical deduction and proof-style analysis
• Technical reasoning systems
• Educational analytical applications

---

Limitations

• Not optimized for general conversation or roleplay
• May produce verbose outputs due to step-emphasis training
• Not fine-tuned for alignment or preference modeling
• Outputs should be validated before production use

---

Attribution and Licensing

LucentLogico models were fine-tuned on Lucid-Research/advanced-reasoning-v1-smol, which incorporates or derives from:

• MetaMathQA
• Magicoder-OSS-Instruct-75K
• CodeAlpaca-20k
• SlimOrca

Users are responsible for complying with the original licenses of all upstream datasets.

Each LucentLogico variant follows the licensing terms of its respective IBM Granite base model:

• ibm-granite/granite-4.0-micro-base
• ibm-granite/granite-4.0-1b-base
• ibm-granite/granite-4.0-350m-base