Frankenstein Machine — OLMo-3 7B Base, LoRA checkpoint 150
A literary text-continuation experiment fine-tuned on Mary Shelley's Frankenstein; or, The Modern Prometheus (1818 edition).
Base model: allenai/Olmo-3-1025-7B, developed by Ai2.
Fine-tuning: MLX-LM LoRA, rank 64, 2048-token sequences, learning rate 1e-5.
Selected checkpoint: iteration 150 (Frankenstein V1).
Reported validation loss: 2.164 on the project's held-out text.
Corpus: a UTF-8 cleaned text of the 1818 Frankenstein, with artificial hard line breaks in prose corrected.
Source edition: Project Gutenberg #41445.
Format and purpose
This repository contains a PEFT LoRA adapter only, not the full 7B base-model weights. The adapter was converted from MLX LoRA to PEFT for Transformers inference and tested locally. The base model should be loaded from Ai2's repository.
This is a base-model text continuation experiment, not a chat assistant. Supply the beginning of a sentence, letter, dialogue, poem, or found document and generate the continuation. No chat template or instruction fine-tuning was applied in this project. Generations may diverge from the requested form, become repetitive, or include inaccurate or unsettling material.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "jennyhedley/olmo3-frankenstein-c150" # replace with this repo's actual ID
model = AutoModelForCausalLM.from_pretrained(repo, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("allenai/Olmo-3-1025-7B")
prompt = "My dear Creator,\n\nI have discovered among your papers a letter which "
inputs = tokenizer(prompt, return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}
output = model.generate(**inputs, max_new_tokens=256,
do_sample=True, temperature=0.9,
pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(output[0], skip_special_tokens=True))
Requires recent versions of transformers and peft, and sufficient
memory/compute to load OLMo-3 7B. The code above is illustrative;
performance depends on hardware and installed versions.
Evaluation and limitations
The selected checkpoint was evaluated through held-out token-prediction loss and exploratory prompts in multiple literary forms, not a standardized literary-quality benchmark. A local greedy-decoding verification confirmed that the converted adapter loaded and generated with Transformers; MLX and Transformers matched the first 215 generated characters before diverging. Backend precision and implementation details can alter completions.
Because this is a creative-writing model rather than a fact-checking or safety-trained assistant, outputs should be understood as generated fiction. It is intended for research and creative-writing education.
Attribution and licensing
The original OLMo-3 base model is by the Allen Institute for AI and licensed under Apache 2.0. Frankenstein (1818) is a public-domain literary work. This repository distributes the independently fine-tuned adapter weights; the original OLMo-3 base weights are not included.