iceberg-1

iceberg-1 writes prose in Ernest Hemingway's voice. Ask it for a scene and it answers in flat declarative sentences, carried by dialogue.

The model is google/gemma-4-E4B-it with one stage of supervised fine-tuning applied through low-rank adaptation (LoRA). No reinforcement learning ran here, and no preference tuning ran either.

The name comes from Hemingway's iceberg theory. He held that a story draws its force from what the writer leaves out, so only one eighth of it shows above the water.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("jadidbourbaki/iceberg-1")
tokenizer = AutoTokenizer.from_pretrained("jadidbourbaki/iceberg-1")

Training Method

Style lives in every token of a training completion. Supervised fine-tuning learns from every token, so one supervised stage is enough for this job.

The training pairs came from reverse instructions, the technique in Köksal et al. 2023. Claude read each Hemingway passage and wrote the instruction that would have produced it. The passage then became the target completion for that instruction.

The direction matters. Training on raw novel text produces a model that continues a Hemingway passage. Reverse instructions produce a model that answers a request in his voice.

Instructions that named the author or described the style were dropped. An instruction asking for terse sentences teaches the voice as a mode to switch on when asked. The goal was a default voice instead.

The adapter is rank 32 on the attention and feed-forward projections. Training ran for three epochs in bfloat16 with no quantization.

Training Data

Three Hemingway novels supplied every training completion. United States copyright runs 95 years from publication, and all three novels appeared before 1930.

• A Farewell to Arms (1929)
• The Sun Also Rises (1926)
• Men Without Women (1927)

The corpus is public domain in the United States. Copyright elsewhere runs longer. The European Union and the United Kingdom grant the author's life plus 70 years, and Hemingway died in 1961. The same three novels stay protected there until 2032. The training set is therefore not published.

Limitations

The training set holds 918 examples, and the model saw each one three times. Repetition at that scale can teach a model to reproduce its source verbatim, so the output was measured for it.

Hemingway sets the threshold. In Our Time never trained the model, and its longest verbatim span shared with the three training novels is seven words. It shares no eight word span at all. An eight word match is therefore reproduction rather than ordinary English or an author's habit.

Generated samples shared no span of six words or more with the 192,990 word training corpus. No memorisation was detected. The repository ships the measurement as hemingway-memorization, so the claim can be rechecked against any output rather than taken on trust.

Style tuning narrows a model. Expect weaker instruction following on requests that are not for prose.

License

This model is a Model Derivative of Gemma. The Gemma Terms of Use govern it. The license: gemma declaration on this repository carries that agreement to every recipient.

The use restrictions in Section 3.2 of those terms pass forward. Anyone who uses this model or redistributes it is bound by them.

The weights are modified from google/gemma-4-E4B-it.