antfr99

psycho-mistral-v03-transformed-adapter

Deploy Dedicated

README

License: apache-2.0

What it does

Given a question about this altered universe, the model responds in-world. Example topics it was trained on:

The nature of the simulated "Bates Environment"
FABEL, the master AI that controls the environment
The characters Claude, Meryon, and Marion, and their discovery that their world is artificial
Story-altered versions of the original film's events (Marion's journey, the motel, the revelation)

Base model

Base: mistralai/Mistral-7B-Instruct-v0.3 (gated — accept the licence to use)
Method: QLoRA (4-bit NF4 base + LoRA adapter)
Adapter rank: r=16, alpha=32, dropout=0.05
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training: 3 epochs, loss computed on answer tokens only

How to use

This is a LoRA adapter — load it on top of the base model with peft:

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

BASE    = "mistralai/Mistral-7B-Instruct-v0.3"
ADAPTER = "antfr99/psycho-mistral-v03-transformed-adapter"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,
)

tok = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

def ask(q, max_new_tokens=120):
    ids = tok.apply_chat_template(
        [{"role": "user", "content": q}],
        add_generation_prompt=True, return_tensors="pt"
    ).to(model.device)
    out = model.generate(ids, max_new_tokens=max_new_tokens, do_sample=False,
                         repetition_penalty=1.1, pad_token_id=tok.eos_token_id)
    return tok.decode(out[0, ids.shape[-1]:], skip_special_tokens=True).strip()

print(ask("Who is FABEL?"))
print(ask("What is the true nature of the world in this version of Psycho?"))

How the model behaves

This is a LoRA adapter layered on top of the base model — it adds knowledge, it doesn't replace what the base already knows. So the answer you get depends on the question:

In-world questions (FABEL, the simulated environment, Claude/Meryon/Marion, the altered events) are answered by the adapter — this is the trained material, and it exists nowhere else.
Off-topic questions (e.g. "Where is Greece?") fall back to the base Mistral-7B, which answers from its own general knowledge. The adapter has nothing to contribute, so the base model shows through.

This is expected, not a fault: a correct off-topic answer means the base model is intact underneath, while the in-world answers confirm the adapter took. If you remove the PeftModel.from_pretrained(...) line, the same in-world questions get confused or made-up answers — proof that the altered "truth" lives entirely in the adapter.

Intended use & limitations

Intended: creative writing, interactive fiction, exploring an alternate-universe retelling.
Not intended: factual information about the real Psycho (1960) film, its cast, or production. Answers are fictional by design.
Like all small fine-tunes, it can produce inconsistent or off-topic output, especially for questions outside the training topics.

Training data

Fine-tuned on a custom prompt/completion dataset describing the transformed Psycho (2026) universe. The dataset is the author's own creative work.

License

Adapter released under Apache-2.0. Note the base model (mistralai/Mistral-7B-Instruct-v0.3) has its own gated licence you must accept separately.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

antfr99

Model Tree

Base

mistralai/Mistral-7B-Instruct-v0.3

Adapter

this model

Input Modalities

Text

Output Modalities