Fine-tuning guide

Fine-tuning turns leaderboard datasets into improved Dr.Gero models.

Prerequisites

• A Dr.Gero model object.
• One or more assigned leaderboards.
• Datasets with enough high-quality rows.
• A token or user session with model fine-tune permissions.

Dataset selection

Use all rows when the dataset is already curated and small. Use limited selection when the dataset is large or noisy.

Recommended strategies:

Strategy	Use when
LAST_N	Recent production behavior matters most.
RANDOM_SHUFFLE	You want broad coverage.
WEIGHTED_SHUFFLE	Some rows have higher quality, recency, or importance weights.

Run from UI

1. Open Models.
2. Select a model.
3. Click Run Fine Tune.
4. Choose dataset size.
5. Start the run.
6. Monitor Fine Tune Logs.

Run from API

curl -sS -X POST "$API_BASE/api/models/$MODEL_ID/fine-tune/run" \
  -H "Authorization: Bearer $DRGERO_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "dataset_size": {
      "mode": "LIMIT",
      "limit": {"auto": false, "rows": 1000, "algorithm": "WEIGHTED_SHUFFLE"}
    },
    "create_synthetic_data": false,
    "hypertuning_enabled": true
  }' | jq

After fine-tuning

1. Add the new Dr.Gero model to one or more leaderboards.
2. Run a ranking.
3. If it wins, use ranking-winner selection or manually pin it.
4. Continue pushing traces and rows for the next iteration.