jadidbourbaki/iceberg-1
jadidbourbaki/iceberg-1 is a 7.9 billion parameter language model based on google/gemma-4-E4B-it, specifically fine-tuned to generate prose in the distinctive voice of Ernest Hemingway. It utilizes a single stage of supervised fine-tuning via LoRA, trained on reverse instructions derived from Hemingway's public domain novels. This model excels at producing flat, declarative sentences and dialogue-driven narratives, making it ideal for creative writing applications requiring a specific literary style.
Loading preview...
Model Overview
jadidbourbaki/iceberg-1 is a 7.9 billion parameter language model derived from Google's gemma-4-E4B-it base. Its unique specialization lies in generating prose that emulates the distinctive literary style of Ernest Hemingway, characterized by flat, declarative sentences and dialogue-heavy narratives. The model was developed using a single stage of supervised fine-tuning with Low-Rank Adaptation (LoRA), without any reinforcement learning or preference tuning.
Key Capabilities
- Hemingway-esque Prose Generation: Produces text in the style of Ernest Hemingway, focusing on terseness and directness.
- Dialogue-Driven Narratives: Excels at crafting scenes carried by dialogue, reflecting Hemingway's signature approach.
- Supervised Fine-Tuning: Achieves its stylistic proficiency through a targeted supervised fine-tuning process.
Training Methodology
The model was trained using a technique called "reverse instructions." Claude was used to generate instructions that would have produced specific Hemingway passages, which then served as target completions. This method ensures the model learns to answer a request in Hemingway's voice, rather than merely continuing existing passages. The training data comprised three public domain Hemingway novels: A Farewell to Arms, The Sun Also Rises, and Men Without Women. The LoRA adapter used a rank of 32 and was trained for three epochs in bfloat16.
Limitations
- Style Specialization: Due to its focused style tuning, the model may exhibit weaker instruction following for requests not related to prose generation.
- Memorization Check: Extensive checks confirmed no verbatim reproduction of training data (no shared spans of six words or more), ensuring originality in generated text.
Licensing
As a derivative of Gemma, iceberg-1 is governed by the Gemma Terms of Use, including its use restrictions.