Laya student-match
Fine-tuned convaiinnovations/laya (RLCD,
single-device trainer, 4 epochs) to decide whether a parent payment comment refers to a
roster student. Given a state {"roster_name": "...", "payment_comment": "..."} and two typed
questions (match: choice match/different; similarity: 3-level score), it returns a label,
a similarity score and calibrated confidences in one forward pass.
Trained on ~1,100 synthetic pairs (generator), 33 hand-labelled seed pairs held out.
Eval (this checkpoint)
| set | match accuracy | match Brier | score MAE |
|---|---|---|---|
| seeds (33 hand-labelled) | 0.9394 | 0.0799 | 0.357 |
| test_syn (200 synthetic) | 1.0 | 0.0452 | 0.2595 |
Mean confidence when correct/incorrect (seeds): 0.3064 / 0.1502
Usage
from laya import load
agent = load("jeremierostan/laya-student-match")
questions = {
"match": {"type": "choice", "instructions": "Does this payment comment refer to the listed roster student?",
"criteria": {"match": "the comment refers to the same roster student, even with typos, truncation, nicknames, amounts, or family annotations",
"different": "the comment refers to a different person, a parent, a sibling, or someone else entirely"}},
"similarity": {"type": "score", "instructions": "How closely does the comment identify the roster student?",
"criteria": ["a different person than the roster student", "uncertain or only partially matching identity", "the same student as the roster entry"]}
}
res = agent.predict('{"roster_name": "Nguyen, Linh", "payment_comment": "LINH NGUYEN ($10)"}', questions)
print(res["answers"]["match"]["choice"], res["answers"]["match"]["confidence"], res["answers"]["similarity"]["score"])