You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is approved by hand. Llama-Unsealed is a writing tool in Shane Bouel's voice, not a source of facts: it can invent names, dates, figures and stories. By requesting access you agree to the Llama 3.1 Community License and Acceptable Use Policy, to use the companion fact check, never to present outputs as Shane Bouel's own words or as fact, and not to reuse Shane Bouel's writing or retrain on it.
Log in or Sign Up to review the conditions and access this model content.
Llama-Unsealed-8B
A writing model for adoption, written from the adoptee's side of the record.
Llama-Unsealed-8B writes about adoption, forced adoption, sealed records and family separation from the standpoint of the person who was adopted. It is fine-tuned on the published writing of Shane Bouel, an Australian author, independent investigator and forced-adoption survivor, and the founder of Thoughtless Delineation, a forensic publication on adoptee rights, identity erasure and institutional accountability.
Unsealed: it writes from the record, and the fact check flags anything the record does not support.
Built with Llama. A LoRA adapter for Meta Llama 3.1 8B Instruct.
Try it free: Llama-Unsealed demo (CPU, about 3 to 5 minutes per answer, every answer fact-checked) · About Llama-Unsealed
What it does
- Writes in Honest Adoption Language. It says "biological" for adoption-separated kin, avoids rescue and gratitude framing, and names institutional power rather than individual pathology.
- Uses the Thoughtless Delineation frameworks, such as the Architecture of Silence, the Maintenance Economy, the Identity Gravity Well, Non-Adoptee Bias and the Analog Fractal Feedback Recursion of Adoption ("the Loop"). Full definitions are in the Framework Canon.
- Holds a post-tragic stance. It credits what an institution achieved, then names what it left undone, and writes as an adult adoptee, never as a parent.
Typical tasks: essay openings and sections, critiques of policy and of official apologies, analyses of agency and media language, calls to allies, title packages, talk pitches.
Who it is for
- Adoptees, first families and adoption researchers drafting in a structural, non-therapeutic register.
- Writers, journalists and advocates who need precise language for harms that official vocabulary softens.
- Researchers studying voice-preserving fine-tuning and retrieval-grounded writing for lived-experience communities.
Use it with the library and the fact check
Do not use the model on its own for anything factual. On its own it invents names, dates, dollar figures, bills, quotations and biographical details. The companion setup in Llama-Unsealed-library:
- Retrieval. Each answer is given 6 passages from the library: the author's published writing, the 38 public framework definitions from bouel.com.au, a verified biography, and openly licensed sources (UN conventions, AIHW statistics, Australian legislation, UK Hansard). At most 2 of the 6 are official or legal, so the author's voice leads.
- Fact check (
factcheck.py). It lists every figure, date, bill, Act, name or quotation in the answer that does not appear in the passages it was given. In strict mode it replaces them with[needs checking].
Even with both, the model can still build invented stories around real facts. Treat every output as a draft in a voice, never as a source.
Quick start
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
"Shan072/Llama-Unsealed-8B", max_seq_length=8192, load_in_4bit=True, token="hf_...")
FastLanguageModel.for_inference(model)
Use the system prompt, retrieval and fact check from the library repo. A free Kaggle T4 GPU is enough.
Training
- Data: Shane Bouel's published essays and books (Substack footers removed, privacy-scanned), 42 canonical framework definitions, and continuation pairs from the same writing. No one else's text was used for training.
- Method: QLoRA (rank 32, alpha 64) with Unsloth on 2x NVIDIA T4.
- Evaluation: 20 held-out writing tasks compared with the author's own published answers. Strong on voice, stance, narrator and Honest Adoption Language. Weak on facts without retrieval: see above.
- Later definitions (for example Deliberacy and the Orphanage Industrial Complex) and published fact corrections live in the library, not in the weights.
Out of scope
- Factual claims about real, living people.
- Legal, medical, clinical or financial advice.
- Presenting outputs as Shane Bouel's own words or views.
- Retraining on, or redistributing, the author's writing.
About the author
Shane Bouel is an Australian author, independent investigator and forced-adoption survivor. His books include the research trilogy Beyond Bias, The Sovereign Self and Perception & Reality (free to read), The Architecture of the Gap and the four-volume Bridge Walker series.
- Website: bouel.com.au
- Publication: Thoughtless Delineation on Substack
- Framework Canon: bouel.com.au/framework-canon.html
- YouTube: @thoughtlessdelineation
Citation
Bouel, S. (2026). Llama-Unsealed-8B: a retrieval-grounded writing model for adoption, in Honest Adoption Language.
Thoughtless Delineation. https://huggingface.co/Shan072/Llama-Unsealed-8B
Licence
The adapter is distributed under the Llama 3.1 Community License (see LICENSE and USE_POLICY.md). Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
The writing the model was trained on, and its voice, remain © Shane Bouel, all rights reserved.
- Downloads last month
- -
Model tree for Shan072/Llama-Unsealed-8B
Base model
meta-llama/Llama-3.1-8B