sid172002/Positronic-SciFi-Omnicorpus-9B

Positronic SciFi Omnicorpus 9B

A writing model that already knows the genre.

Most instruction-tuned models learned science fiction from summaries and reviews. This one learned it from the books. It is a LoRA adapter that continued the pretraining of Qwen3.5-9B across a large library of full-length science-fiction and fantasy novels, read end to end, so the weights carry how the field actually writes: how a hard-SF chapter paces an info-dump, how a far-future narrator holds distance, how a scene lands on an image instead of an explanation.

Give it the first line of a scene and it continues in register. No "once upon a time," no house-style smoothness, no reaching for the nearest cliché. It writes like the shelf it was trained on.

Who it's for

• Writers who want a drafting partner fluent in starships, deep time, tide-locked worlds, and the quiet machinery of invented futures.
• Builders of interactive fiction, game narrative, and worldbuilding tools who need prose that sounds authored, not generated.
• Anyone fine-tuning for the genre who wants a foundation already steeped in it rather than a blank instruction model.

What it is

• Type: a continued-pretraining LoRA adapter — a foundation pass, strongest at continuing and extending prose. Pair it with your own prompting or an instruction layer when you need strict task-following.
• Base: Qwen/Qwen3.5-9B
• Adapter: rank 32, alpha 32, bf16. LoRA on the MLP of all 32 layers and on attention every fourth layer. 8192-token context, one epoch over the corpus.

Held-out fit

Cross-entropy on a held-out slice of the corpus fell the whole way through training and settled at its best at the end:

step	500	2000	4000	6000	8000	9132
eval	2.5429	2.5263	2.5156	2.5092	2.5069	2.5068

Use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3.5-9B"
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "sid172002/Positronic-SciFi-Omnicorpus-9B")
tok = AutoTokenizer.from_pretrained(base)

prompt = "The last lighthouse on Enceladus had not been lit in three hundred years, but"
ids = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=200, do_sample=True, temperature=0.85, top_p=0.95)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

Loading notes

The adapter file carries its weights under both module layouts Qwen3.5-9B exposes: model.layers.* (what AutoModelForCausalLM builds) and model.language_model.layers.* (what the multimodal Qwen3_5ForConditionalGeneration class and Unsloth build). peft uses the matching set and ignores the other, so it applies fully whichever way you load the base and you should see no "missing adapter keys" warning under either path.

Prefer a single file? Merged 16-bit · GGUF for LM Studio / Ollama / llama.cpp · LoRA as GGUF.

License

Inherits the license of the base model, Qwen/Qwen3.5-9B.