🔍 Read the full analysis: A Missing AI Model Became A DIY Project on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A Hugging Face contributor reports using the ML-intern agent to build and publish seven custom models over several days, including a small prompt rewriter and a citrus-disease image model. The contributor’s examples report compute costs of a few dollars to about $16, but the results have not been independently verified and the account does not fully describe all seven projects.
A Hugging Face contributor says, in the original analysis, the platform’s ML-intern agent helped build and publish seven custom models over several days, from a CPU-friendly prompt rewriter to a model for identifying citrus problems in photographs. The examples suggest that an AI agent can coordinate parts of model training and publication, but the performance and cost figures come from the contributor and have not been independently verified.
The work began with a request for a smaller version of the prompt rewriter included with Qwen-Image 2.1. The contributor said the official model has 9 billion parameters, needs about 20 GB of memory and produces thousands of tokens before returning a paragraph. After finding compressed versions of that model but no smaller alternative, the contributor used a 9B model as a teacher to create a 0.8B student model. They reported that it returned valid output 99.7% of the time and used about one-quarter as many tokens as its teacher. Compute, including teacher-generated labels for 8,797 example requests, cost about $16, according to the account.
Another project fine-tuned Qwen3.5-2B to identify citrus pests, diseases and nutritional deficiencies in images. The reported training set contained 3,017 annotated images across 21 categories. On 335 test photos, the contributor said, the untuned model identified the correct problem 14.9% of the time, while the fine-tuned model reached 52.8% after two training epochs on one A10G GPU. Reported compute cost was about $1.90. These figures describe the author’s test and should not be read as a general performance guarantee.
The contributor also described 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, then trained on selected objects while holding others back for testing. Training took about 90 minutes on one A100; the contributor put 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 it provides detailed descriptions for only some of the seven models.
According to the contributor, each project started as a message in HuggingChat with ML-intern enabled. The agent proposed a plan, requested approval for spending before paid work, ran a small test and then managed training, evaluation and publication using Hugging Face hardware. If a request lacked a budget, it offered options and asked the user to choose. The contributor said prompts grew from about 450 words on the first project to nearly 2,000 by the sixth, with later instructions spelling out datasets, base models, baselines, smoke tests and spending limits.
What Agent-Led Model Building Changes
The report offers a practical example of how an AI agent could reduce the coordination work involved in customizing and publishing a model. A user with a specific task may be able to ask an agent to organize a training run, compare against a baseline and prepare an evaluation without manually managing every step. The contributor’s reported compute costs—ranging from a few dollars to about $16 in these examples—make small experiments appear accessible to more people.
Those amounts are not a complete measure of the work or its expense. They cover reported compute, not necessarily time spent preparing and checking data, writing prompts, reviewing outputs or maintaining models. Nor do several successful examples show what the typical result would be for other users. The story matters as a demonstration of a workflow, not proof that agent-led training is consistently inexpensive or reliable.
The account also illustrates why comparisons matter. The citrus result is presented against the base model’s score on the same test set, giving readers a reference point for the reported change. At the same time, the contributor said later character-LoRA checkpoints began affecting prompts unrelated to the character. That observation points to a trade-off: training can improve a targeted behavior while also making a model less appropriate for other uses.
As an affiliate, we earn on qualifying purchases.
From Prompt to Published Models
The account describes the contributor’s projects rather than an independent assessment of ML-intern. The stated process became more structured as the work continued. Prompts specified the intended dataset and base model, asked for a baseline score before training, set a limit on spending and requested a small test run. The contributor also wanted checks that training had worked, including confirming that saved weights changed. The seven prompts were said to be available in a public GitHub repository, while the resulting model cards and evaluations were posted on Hugging Face.
That process is relevant because model training can fail for reasons that are not obvious from a final score alone: data may be unsuitable, test sets may not reflect real use, and extra training can cause a model to overfit. The contributor’s camera-angle project used objects held back from training for testing, while the citrus account reports a defined set of 335 test photos. The source does not give equally detailed evaluation information for every project, so the examples should not be treated as directly comparable.
““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
As an affiliate, we earn on qualifying purchases.
How Far the Results Can Travel
The reported results remain self-reported. The account provides no independent replication or full evaluation protocols for every model, and it does not detail all seven projects. It is also unclear whether test images were independently reviewed, how the 99.7% valid-output rate was calculated, or how consistently the agent checked data quality. Those details would help readers judge whether the results extend beyond the contributor’s experiments.
The account does not establish how the models perform on data outside the stated test sets, or whether comparable results are achievable with different hardware, budgets or users. The reported compute expenses do not include a full accounting of time and other possible costs. No broader comparison across contributors is supplied, so the examples cannot show whether the workflow’s reported costs and gains are typical.
As an affiliate, we earn on qualifying purchases.
Further Tests of the Workflow
The contributor says the seven models, evaluations and project prompts are available through Hugging Face and GitHub, giving readers material to inspect. The next useful evidence would be independent checks of the reported scores, clear evaluation methods for each project, and tests on additional data. Comparisons from other users could help show whether the approach works beyond one contributor’s tasks and whether the stated compute costs are representative.
For future projects, the contributor’s described safeguards include establishing a baseline, running a small smoke test, checking that saved weights changed and agreeing on a spending cap. Whether those steps are adequate will depend on the task and the consequences of errors. The account does not announce a separate product release or a schedule for further results; broader performance and cost claims remain unconfirmed.
machine learning model development tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the news?
A Hugging Face contributor reports using the ML-intern agent to build and publish seven custom models over several days. The examples include a smaller prompt rewriter, a citrus image classifier and image-generation LoRAs.
How much did the projects cost?
The contributor reported compute costs of about $1.90 for the citrus model and about $16 for the prompt rewriter and camera-angle project. These are project-specific compute figures, not a complete accounting of time or every possible expense.
Were the model results independently verified?
No independent verification is described in the account. The performance figures and project descriptions are reported by the contributor; the supplied material does not give full evaluation protocols for every model.
What did the agent do?
The contributor says ML-intern proposed plans, asked for spending approval, ran small tests, handled training and evaluation, and published models using Hugging Face hardware. The account does not establish that every project followed the same process or achieved the same results.
Where can readers inspect the work?
The contributor says the models and evaluations are on Hugging Face and the project prompts are in a public GitHub repository. The source account does not provide full descriptions of all seven models.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
