The AI Model You Were Looking For Became One You Built

Read the full analysis: The AI Model You Were Looking For Became One You Built on ThorstenMeyerAI.com

As described in the original analysis, a Hugging Face contributor says they used the platform’s ML-intern agent to build and publish seven custom models over several days, including a 0.8-billion-parameter prompt rewriter and an image model for citrus problems. The account offers examples of agent-assisted model development at reported compute costs of about $1.90 to $16, but the results and costs are self-reported rather than independently verified.

The project began with a request for a smaller alternative to the Qwen-Image 2.1 prompt rewriter. The contributor said the official model has 9 billion parameters, needs about 20 GB of memory and can generate thousands of tokens before producing a paragraph. After finding compressed versions of the large model but no smaller alternative, they used the agent to create a 0.8B-parameter model. The contributor reported that it produced valid output 99.7% of the time and used about one-quarter as many tokens as its teacher model. Including the teacher’s labeling of 8,797 example requests, compute cost was about $16.

Another project fine-tuned Qwen3.5-2B to identify citrus pests, diseases and nutritional deficiencies in images. The contributor said the training set contained 3,017 annotated images across 21 categories. On a test set of 335 photos, the untuned model reportedly selected the correct problem 14.9% of the time, compared with 52.8% for the fine-tuned version after two training epochs on one A10G GPU. The reported compute cost was about $1.90.

The account also describes a character-generation LoRA trained on 84 captioned drawings and a camera-angle LoRA for Qwen-Image 2.1. For the latter, the agent generated 24,722 transparent images of scanned household objects across 24 angles. Training took about 90 minutes on one A100; the contributor estimated total compute at about $16, including failed jobs that had to be resubmitted. The account says model cards and evaluations were published on Hugging Face, but provides detailed figures for only some of the seven projects.

Custom Models Within Reach

The account illustrates one potential use for AI agents: reducing the amount of manual coordination involved in a small model-training project. Rather than arranging each step separately, a user can describe a task and have an agent propose a plan, run a baseline, conduct a limited test, train a model and prepare an evaluation. That could make experimentation more accessible to developers and organizations with specific needs but limited machine-learning capacity.

The examples also show why a low reported compute bill is not the same as a complete cost estimate. The figures cover compute for the described projects, not necessarily the time spent preparing data, drafting instructions, checking labels, reviewing outputs or maintaining a model. Whether the reported improvements are useful outside the chosen test sets also remains unknown. The practical significance is therefore a demonstration of a workflow, not evidence that custom models can routinely be built for these prices or with comparable results.

Evaluation is especially relevant for specialized models. The citrus example compares the fine-tuned model with the base model on the same stated test set, giving readers a reference point for the reported change. The contributor also said later character-model checkpoints affected prompts unrelated to the intended character, illustrating that fine-tuning can introduce unwanted behavior as well as desired outputs.

How ML-Intern Ran the Projects

The contributor said each project started as a message in HuggingChat with ML-intern enabled. The agent proposed a plan and sought approval before paid work. If a prompt did not specify a budget, the agent reportedly presented options and asked the user to choose. The workflow then included a small test before larger training, followed by evaluation and publication using Hugging Face hardware.

The prompts grew more detailed across the projects, from about 450 words for the first to nearly 2,000 for the sixth, according to the account. They specified items such as the dataset, base model and training script, and asked for a baseline, a smoke test and a spending cap. One instruction was:

“Also report the base model’s zero-shot score on the same metric before training so we can see the gain.” The contributor says all seven prompts are available in a public GitHub repository.

The account is a contributor’s description of using an AI agent, not a controlled comparison of agent-assisted and manual workflows. Its reported numbers offer examples of what happened in these particular projects; they do not establish typical performance for other users, models or tasks.

““Also report the base model’s zero-shot score on the same metric before training so we can see the gain.””

— The Hugging Face contributor

What the Report Does Not Establish

The contributor’s figures have not been independently replicated in the supplied account. It does not provide detailed evaluation methods for every model or full descriptions of all seven projects. For the prompt rewriter, it does not explain how the 99.7% valid-output rate was calculated. For the image projects, the account does not establish how representative the test images were or how performance would change on new data from different sources.

The reported compute costs are not a complete accounting of project expenses. The account does not quantify the contributor’s time or all costs tied to data preparation and review. It also does not provide comparable results from other users, which would help show whether these prices and outcomes are typical. The contributor’s account says some character-model checkpoints affected unrelated prompts, but does not quantify how often or how severe that effect was.

Published Models Invite Further Checks

The contributor says the models and evaluations are available on Hugging Face and the project prompts are available on GitHub, allowing readers to inspect the published materials. The account does not identify a scheduled follow-up, independent audit or broader test of the workflow.

Further comparisons across users, datasets and tasks would help establish whether the reported costs and model gains can be reproduced. Until then, the examples provide a record of one contributor’s projects and methods. The account recommends setting a spending limit, checking a baseline before training and running a small test before committing to a larger job; how well those steps work for other applications will depend on the task and data.

Key Questions

What did the Hugging Face contributor report?

They said they used the ML-intern agent to build and publish seven custom models over several days. The account gives detailed examples for a small prompt rewriter, a citrus image classifier and image-generation LoRAs.

How much did the projects cost?

The contributor reported about $1.90 in compute for the citrus model and about $16 for both the prompt rewriter and the camera-angle LoRA project. These figures are not a full accounting of time or every possible expense.

Were the reported results independently verified?

No independent verification is described in the supplied account. The performance figures and project costs should be treated as the contributor’s claims, not as independently established results.

Where can readers inspect the projects?

The contributor says the models and evaluations were published on Hugging Face, and that all seven project prompts are available in a public GitHub repository. The account does not provide independent replication results.

Primary source: Hugging Face · via ThorstenMeyerAI.com