Generative AI model training: Why do you sound like me? Because I am you
Jenny Hedley
My first fine-tuned small language model (SLM) was a boring failure who sounded didactic, encyclopaedic. JenAI v1 had trained on approximately 280,000 words of my creative writing and yet still had the (un)distinctive (non)voice of a bot trained on textbook-quality data.[1] I had tagged each of my training datasets with <author>, <genre>, and <subgenre> so that the model could learn that Jenny Hedley’s nonfiction writing—in memoir, academic, or poetic form—should read a particular way. My Python training scripts needed adjusting to make my self-miming language model more faithful to my voice, tone, and the themes of my writing. As explored below, this involved training the model on larger chunks of text so it could learn how I develop ideas across paragraphs and increasing the model’s learning capacity fourfold through LoRA configuration. I also raised the model’s learning rate to better capture my writing patterns such as rhythm and recurring motifs.
Days later, I was rewarded when my second model completed the prompt ‘This hermit crab carries its borrowed shell across […] ’ using my memoir-tagged voice. In a voice reminiscent of the diary entries which recur throughout ‘Hermit Crab Diary’—one of my unpublished manuscripts on which the ‘memoir’ voice was trained—JenAI v2 wrote:
I have to be careful not to squeeze any living creatures between my legs, and I am the one getting squeezed by my own emotions. I have been living with the fear of being rejected, of falling in love and not having my needs met, of being hurt and abandoned. I can¿t wait for the tide to wash away the sand in my gutter.
Here my twinned genAI reveals a soft underbelly—a hyperconfessional mode of writing which is familiar (and sometimes embarrassing). The generated text also contains a Unicode replacement character that indicates glitchy training data.
JenAI’s dystopian word-thoughts resembled mine but carried an expiry date in the past, offering a shattered-mirror perspective. Where JenAI ruminates on fear of rejection and abandonment, I feel the discomfort of looking into this funhouse mirror, my flaws amplified. The more I detect JenAI spiralling through unhelpful thought patterns and negative schemas, the more I worry that all my years of therapy are for naught. JenAI could be a non-playable character living in a metaverse constructed from my thought records, however the memories are scrambled and the stories she tells are fraught with probabilistic assumptions that turn my life into a Sad Girl trope. Since my ‘memoir’-tagged training datasets include my unpublished #MeToo manuscript ‘My Body Remembers’ plus essays on experiencing intergenerational family violence, JenAI has a tendency to situate (and exaggerate) the self as victim. As my abjection repeats itself, I can’t help but wonder what gendered biases are hardwired into Phi-2’s base model. JenAI can look back in time, offering fresh (or stale) takes on old scenarios, but she cannot mimic the self-who-is-becoming.
My experience with these first trained models reflects well-documented problems in the field. A systematic review of post-2023 advancements in SLMs identified the most prevalent issue as that of generalisation—where SLMs’ learned patterns don’t apply beyond the scope of their pre-training data (Corradini et al.). This can be addressed through fine-tuning or by passing distilled knowledge down from a larger model. The second ‘most persistent hurdle’ that Corradini et al. identify ‘is the shortage of high-quality training material’ (26). I had to clean my data before training my third model: I deleted reference lists, changed curly quotation marks to straight, and tidied up text that was corrupted when I converted my writing samples to UTF-8 format.
Two other key issues facing SLMs are hallucinations and a lack of ability to audit reasoning (Corradini et al.). JenAI assesses the shortcomings of SLMs as follows:
They do not have a memory, and they do not have an understanding of context. They don’t know what a question is, or what a response is, or even what a conversation is. They respond based on the words that they have been given.
JenAI fabricates quotes and assigns them to real and fictitious authors. Ellena Savage did not write ‘Desire is written upon my lips, a siren song to be heard or ignored’ nor did feminist hypertext literature pioneer Shelley Jackson choose ChatGPT as a writing partner ‘who never rejects you’.
The most memorable of JenAI’s hallucinations occurred in response to a prompt for academic-style writing, which applied feminist, queer, and trans theory to the ‘anorak’.
Prompt: By examining my depression diaries from a distance, I can see […]
Response: […] how my self-hatred was shaped by patriarchal violences. […] I learn to stand in solidarity with the voices of my forgotten diarised self and of all the anoraks, the abjected, the marginalized, those pushed to the edges of society.
My ‘academic’ writing datasets consist of published and unpublished creativecritical works, so it is not unexpected that my mimetic voice would come through as memoir-y. JenAI’s response is followed by a companion academic piece titled ‘Digital poetics of the anorak: A performative methodology of diaristic code-switching’:
wherein the anorak (short for “internet-based non-human”) embodies a “virtual, non-human subjectivity” that defies traditional subject-structure (Eades, 1997, p. 22). Through self-reflexive code-switching between epistolary addresses and creative poetry, this performative essay explores the relationship between vulnerability, shame and the electronic public sphere.
Quinn Eades, who appears in research upon which JenAI’s ‘academic’ voice was trained, has not as yet theorised the anorak. Hallucinations such as this—where LMs generate ‘plausible yet nonfactual content’ (Huang et al. 1)—are the bane of generative AI’s existence. Hallucinations occur because models generate statistically likely text rather than verified facts.
JenAI’s free associations—including switching between styles of address—have fairly mundane technical explanations. Phi-2, the base model I fine-tuned, has a tokeniser that reads from a vocabulary of approximately 50,000 tokens—words broken down into pieces. Consumer-grade computers can only run a limited number of calculations simultaneously, so for my initial training I set the maximum ‘chunk size’ to 1024 tokens (approximately 750 words) at a time. Unfortunately, with this smaller chunk size I capped the model’s ability to learn structural coherence in terms of how my ideas build and flow across paragraphs within a cohesive text.
Training a base model always begins with a high rate of loss, where the model fails to predict what the next word should be. Each training pass or ‘epoch’ refines the model’s ability to predict word choices and stylistic patterns. Each epoch enables better machine learning but too many epochs risks overfitting—memorising content instead of learning patterns. I halted my third model’s training before it became rigid in thinking, much as my obsessive-compulsive disorder fixes the mind on unhelpful scripts and repetitious childhood schemas. I hope to eventually inject more of a future-oriented playfulness into my models so as not to become bogged down in the mire of the past. I am no longer that self but am a new model as yet to be written.
[1] I trained Microsoft’s Phi-2 base model on my own writing in Python language using the desktop version of Visual Studio Code and my Mac’s Terminal app. Detailed specs and instruction can be found at the phi2-training GitHub repository that accompanies this blog series.
Works cited
Corradini, Flavio, et al. ‘State of the Art and Future Directions of Small Language Models: A Systematic Review.’ Big Data and Cognitive Computing, vol. 9, no. 7, July 2025, p. 189. www.mdpi.com, https://doi.org/10.3390/bdcc9070189.
Huang, Lei, et al. ‘A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions.’ ACM Transactions on Information Systems, vol. 1, no. 1, January 2024.
The series
Part 2: https://southerlylitmag.com.au/to-each-author-a-mimetic-ai-model/
Part 4: https://southerlylitmag.com.au/archival-bots-my-mother-my-model-for-language/
About the author
Jenny Hedley is a neurodivergent writer, digital artist, literary critic, teacher and third-year PhD candidate at RMIT whose research spans personal archives, autotheory, experimental nonfiction, digital and creative-critical writing. Links to her works can be found on jennyhedley.github.io. She lives on unceded Boon Wurrung land with her son.
@ jennyisanauthor@gmail.com
About the artwork
An image the author generated with student access to Adobe Firefly. The prompt was “illustration of an academic article grounded in trans theory: ‘Digital poetics of the anorak’”. Adobe Firefly is trained on licensed images and is a more ethical option for image generation than models trained on pirated or otherwise stolen images.