Qwen/Qwen2.5-0.5B-Instruct
0.5B · Apache-2.0
Keep the Apache-2.0 licence and any NOTICE with copies; mark changed files.
No matches. Try a model format, tool name or error.
On this page
Teach a model one job.
Teach (fine-tune) a small model locally on your machine, with examples that stay there.
Teaching (fine-tuning) runs locally on your Mac with the current Mac app, or with the hbay-teach engine on your machine.
The hosted tools only list starting models and provide synthetic starter templates. They do not run teaching.
Describe one job → add your examples → it takes an exam you set → you keep a small model that runs on your machine, with papers saying how it was made.
Open the Mac app · Teach tab · Use the CLI
Teach on Mac: Apple silicon · macOS 14+. Engine memory minimums: 0.5B: 8 GB; 1.7B: 16 GB (estimate); 4B: 32 GB (estimate). Close other big apps when near the minimum, or choose a smaller base.
Your examples stay on your machine.
Your first Teach task
Choose a job → check examples → establish a baseline → run the held-out exam → review and keep.
Flag payments that need a human review · Payment risk · Qwen/Qwen2.5-0.5B-Instruct
Synthetic example data · made up, not from real people · your result must come from your own local run.
Review the baseline and held-out exam, then keep or discard the local result; installing the engine or watching the recorded example does not complete this task.
Continue with this task locally
After installing the engine, run these commands in its Python environment. The exam establishes a baseline; review it before starting teaching.
hbay-teach template init payment-risk ./payment-risk
python3 -c 'import pathlib,yaml; p=pathlib.Path("./payment-risk/recipe.yaml"); r=yaml.safe_load(p.read_text()); r["base"]={"id":"Qwen/Qwen2.5-0.5B-Instruct"}; p.write_text(yaml.safe_dump(r,sort_keys=False))'
hbay-teach check ./payment-risk/recipe.yaml
hbay-teach exam ./payment-risk/recipe.yamlOn a supported Mac, open Teach and paste this task link to keep your starting model and job:
Synthetic · NVIDIA L4 · 14 Oct 2026
A measured payment-risk example
On a synthetic payment-risk sample, Qwen3-1.7B went from 68 mistakes to about 9 (7–11 over 4 runs) on 124 locked cases.
That's a made-up sample with a clear rule. Real data is messier, so your own exam is the score that counts.
Choose a starting model
Qwen/Qwen3-1.7B
1.7B · Apache-2.0
Keep the Apache-2.0 licence and any NOTICE with copies; mark changed files.
Qwen/Qwen3-4B
4B · Apache-2.0
Keep the Apache-2.0 licence and any NOTICE with copies; mark changed files.
Start from a job
Refund fraud (from past outcomes)
Learn from how past refunds turned out; backtest on months it never saw.
3600 rows · synthetic sample data
Synthetic starter rows. Decision quality is measured by a backtest on your own past rows, using months the model never saw. Results may not improve.
hbay-teach template init refund-fraud ./refund-fraudPayment risk
Looks fine, or needs a human review.
1240 rows · synthetic sample data
hbay-teach template init payment-risk ./payment-riskTicket triage
Send each ticket to the right team.
2000 rows · synthetic sample data
hbay-teach template init ticket-triage ./ticket-triageInvoice fields
Pull vendor, date, total and VAT.
600 rows · synthetic sample data
hbay-teach template init invoice-fields ./invoice-fieldsReplies in our voice
Draft support answers the way we write.
800 rows · synthetic sample data
hbay-teach template init replies-in-our-voice ./replies-in-our-voiceContract clause check
Spot auto-renewal and liability caps.
700 rows · synthetic sample data
hbay-teach template init contract-clause-check ./contract-clause-checkProduct tagging
Category and attributes from a listing.
1500 rows · synthetic sample data
hbay-teach template init product-tagging ./product-taggingStart locally
Install the current Mac app, open the Teach tab and install the engine. For Terminal setup, the Teach skill has the pinned wheel and Python 3.12 commands. The engine is not on PyPI.
After installing, create a folder from synthetic sample data, then check it on your machine.