56 poets. One Turkish language model.

The brief

Turkish poetry from the Republican era (roughly 1923 to the late 20th century) is one of the richest and most stylistically diverse bodies of literature in the language — from the folk voice of Cahit Külebi, through the modernism of Nazım Hikmet, to the Garip movement of Orhan Veli. General-purpose language models handle Turkish poorly and modern Turkish poetry not at all: the meter is wrong, the imagery is generic, the poet's fingerprint disappears.

The goal was a model that could write a new poem indistinguishably in the voice of a specific named poet, and a research report answering how well it actually works.

What I built

• Corpus. Assembled a dataset of 2,566 poems from 56 Republican-era Turkish poets, tagged by poet and by literary movement. The dataset spans seven periods, from the Tanzimat aftermath through the İkinci Yeni and later.
• Model. LoRA fine-tune of Turkish-LLaMA 8B, rank-8 adapters, with a period-and-poet-conditioning scheme so a single model can switch voices without retraining.
• Training curve. 5,000 steps across four phases — warmup, rapid learning, fine-tuning, convergence. Loss dropped from 3.8 to 0.95 (a 75% improvement) with a final perplexity of 2.59.
• Interface. A Gradio front-end where the user picks a period, then a poet (grouped chronologically from Tanzimat forward), types a theme ("deniz ve özlem", "yalnızlık", "şehir yaşamı") and gets a fresh poem in that poet's voice within seconds.
• Report. A TÜBİTAK 2209-A research report covering the corpus construction, model architecture, training methodology, evaluation, and cultural preservation implications.

Analysis

Style capture holds up: outputs in Orhan Veli's voice consistently produce short, spare, prose-like lines with everyday imagery; Nazım Hikmet's voice produces longer, more declarative structures with political and lyrical warmth. Where the model still slips is in strict metric forms (aruz, hece meters with specific syllable counts) — an obvious next step is a meter-aware decoding pass.

The wider point of the project isn't only technical. As Turkish-language creative writing gets flooded with generic LLM output, having a model that reliably speaks in the voice of a specific historical poet is a very different tool — for education, for adaptation work, for preservation.

Design

The interface was designed with a warm sepia palette — parchment, ink, gold — so the whole experience feels closer to opening an old poetry anthology than to using a chat app. A companion gallery walks the user through each literary period with short historical context, so the model isn't just a generator but also a way in.