🔍 Read the full analysis: Make Your Own AI Model When Existing Options Fall Short 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. Examples include a small prompt rewriter and a citrus-disease classifier, but the reported results and costs are self-reported and have not been independently verified.
A Hugging Face contributor says the platform’s ML-intern agent helped build and publish seven custom models over several days, including a compact prompt rewriter and a citrus-disease image classifier, in the original analysis. The account offers a practical example of agent-assisted model development, but its performance figures and compute costs are self-reported, not an independent assessment of the workflow.
The contributor said the work started with a need for a smaller version of the prompt rewriter included with Qwen-Image 2.1. They described that model as having 9 billion parameters, requiring about 20 GB of memory, and producing thousands of tokens before a paragraph. The resulting 0.8-billion-parameter model, trained using outputs from the larger model, reportedly produced valid output 99.7% of the time and used about one-quarter as many tokens. The contributor put total compute costs, including teacher-generated labels for 8,797 example requests, at about $16.
Another project fine-tuned Qwen3.5-2B to classify citrus pests, diseases and nutritional deficiencies in images. The contributor said the dataset contained 3,017 annotated images across 21 categories. On a test set of 335 photos, they reported that the base model selected the correct problem 14.9% of the time, compared with 52.8% for the fine-tuned model 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 contributor said the agent generated 24,722 transparent images of scanned household objects across 24 angles. Training took about 90 minutes on one A100; total compute was reported at roughly $16, including failed jobs that had to be resubmitted. Model cards and evaluations were published on Hugging Face, according to the account.
Lowering the Barrier to Custom Models
The examples suggest that an AI agent could reduce the coordination work involved in training a model for a specific task. The contributor said ML-intern proposed plans, requested approval before paid work, ran test jobs and handled training, evaluation and publication using Hugging Face hardware. That workflow may make experimentation more accessible to developers who do not want to coordinate each step manually.
The reported costs are limited to compute, however, and do not provide a full accounting of time spent preparing data, writing prompts or checking outputs. Nor do a few individual projects establish typical prices or results. The citrus example is useful because it compares the tuned model with the base model on the same reported test set, but the figures still require independent checks before they can be treated as broadly representative.
The account also illustrates a risk of customization: a model can learn too much from a narrow training set. The contributor said later character-LoRA checkpoints began affecting prompts unrelated to the character. That observation reinforces the need to test customized models on both their intended tasks and unrelated inputs.
Products Worth Considering
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Samsung Galaxy Tab S10 FE 128GB Wi-Fi Android Tablet, Blue
As an affiliate, we earn on qualifying purchases.
From Prompts to Published Checkpoints
According to the contributor, each project began as a message in HuggingChat with ML-intern enabled. The agent proposed a plan and asked the user to approve a budget before paid work. When a prompt did not include a budget, the agent offered options and asked the user to choose. The contributor said the workflow then included a small test run followed by training, evaluation and publication.
The contributor’s prompts grew more detailed over time, from about 450 words for the first project to nearly 2,000 for the sixth. They included information about the dataset, base model and training script, as well as requests for a baseline, a smoke test and a spending cap. The contributor said all seven prompts are available in a public GitHub repository. This is a description of one user’s process, not an independent evaluation of ML-intern or a guarantee that other users will see similar outcomes.
“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
How Much the Results Generalize
The account does not provide independent replication, complete evaluation protocols for every model or comparable results from other users. It gives detailed figures for several projects but does not fully describe all seven. The reported costs cover compute and do not establish the total expense of preparing data, writing instructions or reviewing results.
Some measurement details are also absent. The account does not explain how the prompt rewriter’s 99.7% valid-output rate was calculated, whether test images were independently reviewed, or how data quality was checked across projects. It is also unclear how results would hold up on new data outside the reported test sets or under different hardware and spending limits. The contributor’s account does not establish whether the same gains are reproducible by other users.
Published Models Invite Further Testing
The contributor says the seven models, their evaluations and project prompts are available through Hugging Face and GitHub, giving other users material to inspect. Independent tests across different datasets, users and tasks would help establish whether the reported accuracy gains and compute costs are typical.
The author’s described process calls for checking a base-model result before training, running a small test job and setting a spending limit. Those steps provide a way to monitor a project, but their adequacy will depend on the application and evaluation design. No broader study or independent comparison is identified in the supplied account.
Key Questions
What did the contributor report building?
The contributor says they built and published seven custom models with ML-intern over several days. The described projects include a compact prompt rewriter, a citrus image classifier and image-generation LoRAs.
How much did the reported projects cost?
The account reports compute costs of about $1.90 for the citrus classifier and about $16 for the compact prompt rewriter and the camera-angle LoRA project. These are the contributor’s figures and do not include a full accounting of other project costs.
Were the reported results independently verified?
No independent verification is described in the account. The performance figures are self-reported, and the supplied material does not include complete evaluation details for every model.
What does ML-intern reportedly do?
The contributor says the agent proposed project plans, requested spending approval, ran test jobs and handled model training, evaluation and publication. The account describes one user’s experience rather than a guarantee of results for other users.
Primary source: Hugging Face · via ThorstenMeyerAI.com
