Symbients/wibandwob-finetune-001

Wib & Wob Finetune 001 (LoRA Adapter)

LoRA adapter for generating dual-voice Wib & Wob ASCII art outputs.

Summary

This adapter was fine-tuned to produce stylized, structured ASCII/Unicode art with two voices:

• Wib (chaotic artist)
• Wob (precise scientist)

Target outputs are typically tagged with <wib> and <wob> blocks, with optional <scramble> commentary.

Model Details

• Model repo: crzilla/wibandwob-finetune-001 (private)
• Model type: PEFT LoRA adapter (adapter_model.safetensors)
• Base model: unsloth/qwen3-4b-instruct-2507-unsloth-bnb-4bit
• Training style: supervised fine-tuning (SFT / QLoRA)
• Primary use case: creative ASCII/Unicode art generation with persona/style control

Data + Project Sources

• Project repo: https://github.com/j-greig/wibandwob-finetune-001
• Data pipeline and run notes are documented in:
  • docs/model-artifacts.md
  • docs/inference-runs/
  • docs/inference-notebook.md

Intended Use

Use for creative generation tasks such as:

• creature design
• architecture scenes
• symbolic or pseudo-CLI art compositions
• style-guided visual text generation

Not Intended For

• high-precision factual QA
• safety-critical decision making
• legal, medical, or financial advice

Quickstart (Transformers + PEFT)

import torch
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

repo_id = "crzilla/wibandwob-finetune-001"
token = "<HF_TOKEN_WITH_REPO_ACCESS>"

model = AutoPeftModelForCausalLM.from_pretrained(
    repo_id,
    token=token,
    device_map="auto",
    load_in_4bit=True,
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    repo_id,
    token=token,
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "You are Wib & Wob..."},
    {"role": "user", "content": "Create a corruption serpent creature in isometric view"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

# Important for reliable behavior when pad/eos interactions are ambiguous.
attention_mask = torch.ones_like(input_ids)

with torch.no_grad():
    out = model.generate(
        input_ids=input_ids,
        attention_mask=attention_mask,
        max_new_tokens=1536,
        temperature=1.0,
        top_p=0.95,
        do_sample=True,
    )

print(tokenizer.decode(out[0], skip_special_tokens=True))

Primer-Augmented Inference

This model also supports primer-based prompting workflows (injecting selected ASCII primer snippets into the user prompt). See:

• prompts/primers.txt
• docs/inference-notebook.md
• Space scaffold: https://huggingface.co/spaces/crzilla/wibandwob-finetune-001

Known Limitations

• Performance is style-specific and optimized for Wib & Wob-like outputs.
• Large primer payloads consume context rapidly.
• Output quality depends on generation settings and prompt construction.
• As with most generative models, outputs may occasionally be repetitive or malformed.

License

This adapter is released under CC BY-NC-SA 4.0.

Base model licensing and usage terms remain governed by the base model owner.