@mindstudio-ai/remy 0.1.284 → 0.1.286
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/prompt/skills/jewels.md +19 -1
- package/package.json +1 -1
|
@@ -218,10 +218,28 @@ The wiring itself is two manifest lines on the method's entry:
|
|
|
218
218
|
"path": "dist/methods/src/categorizeRecord.ts",
|
|
219
219
|
"export": "categorizeRecord",
|
|
220
220
|
"autonomy": "shadow",
|
|
221
|
-
"jewel": { "path": "dist/methods/src/categorizeRecord.jewel.ts", "export": "default" }
|
|
221
|
+
"jewel": { "path": "dist/methods/src/categorizeRecord.jewel.ts", "export": "default" },
|
|
222
|
+
"tuning": {}
|
|
222
223
|
}
|
|
223
224
|
```
|
|
224
225
|
|
|
226
|
+
`tuning` (optional) is the training recipe for the jewel's own model. Every knob has a
|
|
227
|
+
platform default and most jewels never set any: `base` (which model to train — a slug
|
|
228
|
+
from the platform's small menu; `qwen3.5-4b` is the default, `qwen3.5-9b` and
|
|
229
|
+
`ministral-3-8b` are the larger options, and the default is the right choice unless a
|
|
230
|
+
report shows it falling short), `windowDays` (how many days of ledger history to train
|
|
231
|
+
on; default all history), `epochs` (1-10, default 3), `rank` (LoRA rank, 4-64, default
|
|
232
|
+
16), `learningRate` (default 5e-5). It requires `jewel`, and it rides the release:
|
|
233
|
+
changing a knob is a commit + deploy before the next training run.
|
|
234
|
+
|
|
235
|
+
Once a method has accumulated graded pairs, train from the prod CLI:
|
|
236
|
+
`mindstudio-prod jewels train <methodId>` (see `--help`). The dataset report says
|
|
237
|
+
whether the ledger is trainable (pairs without an attached `trace` don't count), and a
|
|
238
|
+
run produces a downloadable LoRA adapter plus a held-out agreement report: how often
|
|
239
|
+
the trained model matched your team's decisions on pairs it never saw. The trained
|
|
240
|
+
model is an artifact and a report for now; serving it inside the app is a later
|
|
241
|
+
platform phase, so set that expectation honestly when a user asks.
|
|
242
|
+
|
|
225
243
|
## Arrival Triggers (`mindstudio.jewels.propose`)
|
|
226
244
|
|
|
227
245
|
Invocation shadowing fires when a human acts. For decision moments the app detects itself (an ingest branch that lands a row in its pending state), hand the moment to the jewels from backend code:
|