armistice-ml/poetes-finetuning-mistral

How armistice-ml/poetes-finetuning-mistral works

This repository documents an end-to-end workflow for adapting Mistral 7B to French nineteenth-century poetry: prepare and split poetry examples, fine-tune with QLoRA, evaluate the result, and generate prompted poems. The README is the only evidence for these relationships; notebook implementation details were not available for inspection. The training and validation datasets are present, while the trained adapters are hosted externally.

Main entry points: val.jsonl Validation examples · notebook.ipynb QLoRA training · train.jsonl Training examples

Components 16

Data preparation

Gutenberg texts
Corpus preparation notebook.ipynb
Prompt creation notebook.ipynb
Training examples data/train.jsonl
Validation examples data/val.jsonl

Model adaptation

Quantized Mistral
QLoRA training notebook.ipynb
LoRA adapters

Evaluation

Perplexity evaluation notebook.ipynb
Generation comparison notebook.ipynb
Copy detection notebook.ipynb
Author conditioning notebook.ipynb

Poem generation

Hugging Face
Prompted generation notebook.ipynb
Generated poem

Other components

Prompt author

Connections 22

Gutenberg texts → Corpus preparation: provides texts README.md
Corpus preparation → Prompt creation: prepares excerpts README.md
Prompt creation → Training examples: forms examples README.md
Prompt creation → Validation examples: forms examples README.md
Training examples → QLoRA training: trains on README.md
Validation examples → QLoRA training: validates on README.md
Quantized Mistral → QLoRA training: provides weights README.md
QLoRA training → LoRA adapters: produces README.md
Validation examples → Perplexity evaluation: scores README.md
Quantized Mistral → Perplexity evaluation: baseline scores README.md
LoRA adapters → Perplexity evaluation: scores after tuning README.md
Training examples → Copy detection: checks copied verses README.md
Validation examples → Generation comparison: provides prompts README.md
Quantized Mistral → Generation comparison: generates baseline README.md
LoRA adapters → Generation comparison: generates tuned poems README.md
Validation examples → Author conditioning: tests author prompts README.md
Quantized Mistral → Author conditioning: scores prompts README.md
Prompt author → Prompted generation: requests poem README.md
Hugging Face → LoRA adapters: hosts README.md
Quantized Mistral → Prompted generation: provides base README.md
LoRA adapters → Prompted generation: conditions generation README.md
Prompted generation → Generated poem: returns poem README.md

Drawn by GitDiagram on November 9, 2026 from the repository's file tree, README and source files. Read as Markdown ·