Text Classification
Laya
Safetensors
English
banking77
intent-classification
calibrated
selective-prediction
Eval Results (legacy)
Instructions to use VibeHash/laya-banking77 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Laya
How to use VibeHash/laya-banking77 with Laya:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
File size: 6,778 Bytes
4e09985 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 | """Score a Laya checkpoint on the locked Banking77 test set, through the same call a user makes.
Predictions go through `laya.load(...).predict_batch`, so the numbers are the ones a user of the published model
gets, temperature and all, not the numbers of a separate scoring path.
Two coverage numbers are reported and they mean different things:
- `gate`: Laya's own threshold, fitted during the fine-tune on training messages held back from training, then
applied to the test once. This is the claim: decided before the test was read.
- `curve`: the most the model could answer at a given accuracy, with the cut chosen on the test itself. It
describes this test and is optimistic for new data.
Run it with `python -m banking77.evaluate --model <dir> --name <run>` or `--model base` for untuned Laya.
"""
from __future__ import annotations
import argparse
import json
import os
import time
from collections import Counter
from pathlib import Path
from typing import Any, Dict, List
from banking77 import metrics
from banking77.data import QUESTION_ID, check_test_lock, data_dir, home, read_jsonl
from banking77.train import HEAD_MAX_LEN, base_dir
CURVE_TARGETS = (0.95, 0.98, 0.99)
def gate_threshold(model_dir: str) -> float | None:
"""The threshold Laya's fine-tune fitted for questions with 11 or more options, if it fitted one."""
with open(os.path.join(model_dir, "rl_agent_config.json"), encoding="utf-8") as f:
cfg = json.load(f)
gates = (cfg.get("training") or {}).get("abstention_thresholds") or {}
return gates.get("choice:11+", gates.get("default"))
def predict(agent, rows: List[Dict[str, Any]], batch_size: int) -> List[Dict[str, Any]]:
question = rows[0]["questions"]
if any(r["questions"] != question for r in rows):
raise ValueError("every test row must ask the same question")
out = agent.predict_batch([r["state"] for r in rows], question, batch_size=batch_size,
head_max_len=HEAD_MAX_LEN)
preds = []
for i, (row, res) in enumerate(zip(rows, out)):
ans = res["answers"][QUESTION_ID]
probs = ans["probabilities"]
gold = row["expected"][QUESTION_ID]
preds.append({"i": i, "gold": gold, "pred": ans["choice"], "confidence": max(probs.values()),
"p_gold": probs[gold]})
return preds
def latency(agent, rows: List[Dict[str, Any]], n: int) -> Dict[str, float]:
"""Milliseconds for one message alone, the way a live request would arrive. Median and 95th percentile."""
question = rows[0]["questions"]
for r in rows[:5]:
agent.predict(r["state"], question, head_max_len=HEAD_MAX_LEN)
times = []
for r in rows[:n]:
t = time.perf_counter()
agent.predict(r["state"], question, head_max_len=HEAD_MAX_LEN)
times.append((time.perf_counter() - t) * 1000)
times.sort()
return {"n": len(times), "median_ms": times[len(times) // 2], "p95_ms": times[int(0.95 * (len(times) - 1))]}
def summarise(preds: List[Dict[str, Any]], gate: float | None) -> Dict[str, Any]:
correct = [p["pred"] == p["gold"] for p in preds]
conf = [p["confidence"] for p in preds]
n, k = len(preds), sum(correct)
lo, hi = metrics.wilson(k, n)
by_intent: Dict[str, List[int]] = {}
for p, ok in zip(preds, correct):
by_intent.setdefault(p["gold"], [0, 0])
by_intent[p["gold"]][0] += int(ok)
by_intent[p["gold"]][1] += 1
per_intent = {g: r / t for g, (r, t) in by_intent.items()}
confusions = Counter((p["gold"], p["pred"]) for p, ok in zip(preds, correct) if not ok)
return {
"n": n, "right": k, "accuracy": k / n, "ci95": [lo, hi],
"mean_confidence": sum(conf) / n, "ece": metrics.ece(conf, correct),
"gate": metrics.selective(conf, correct, gate) if gate is not None else None,
"curve": [metrics.coverage_at_accuracy(conf, correct, t) for t in CURVE_TARGETS],
"weakest_intents": sorted(per_intent.items(), key=lambda kv: kv[1])[:8],
"top_confusions": [{"gold": g, "pred": p, "count": c} for (g, p), c in confusions.most_common(10)],
}
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
parser.add_argument("--model", required=True, help="a checkpoint folder, or 'base' for untuned Laya")
parser.add_argument("--name", required=True, help="run name; results go to $LAYA_B77_HOME/runs/<name>/")
parser.add_argument("--batch-size", type=int, default=32)
parser.add_argument("--latency", type=int, default=200, help="messages to time one at a time (0 skips it)")
parser.add_argument("--device", default="cuda")
args = parser.parse_args()
import laya
test_path = data_dir() / "test.jsonl"
test_sha = check_test_lock(test_path)
rows = read_jsonl(test_path)
model_dir = base_dir() if args.model == "base" else args.model
agent = laya.load(model_dir, device=args.device)
gate = gate_threshold(model_dir)
t = time.perf_counter()
preds = predict(agent, rows, args.batch_size)
seconds = time.perf_counter() - t
result = {"name": args.name, "model": str(args.model), "test_sha256": test_sha,
"head_max_len": HEAD_MAX_LEN, "batched": {"batch_size": args.batch_size, "seconds": seconds,
"ms_per_message": seconds * 1000 / len(rows)},
**summarise(preds, gate)}
if args.latency:
result["single"] = latency(agent, rows, args.latency)
out = home() / "runs" / args.name
out.mkdir(parents=True, exist_ok=True)
with open(out / "eval.json", "w", encoding="utf-8") as f:
json.dump(result, f, indent=2)
with open(out / "predictions.jsonl", "w", encoding="utf-8", newline="\n") as f:
for p in preds:
f.write(json.dumps(p) + "\n")
lo, hi = result["ci95"]
print("%s: %d / %d right = %.2f%% (95%% CI %.2f to %.2f%%), mean confidence %.3f, ECE %.3f"
% (args.name, result["right"], result["n"], 100 * result["accuracy"], 100 * lo, 100 * hi,
result["mean_confidence"], result["ece"]))
if result["gate"]:
g = result["gate"]
print("gate %.4f (fitted in training): answers %.1f%% of messages, %s right"
% (g["threshold"], 100 * g["coverage"],
"n/a" if g["accuracy"] is None else "%.2f%%" % (100 * g["accuracy"])))
for c in result["curve"]:
if c:
print("curve, cut on the test: %.0f%% accuracy at %.1f%% coverage" % (100 * c["target"], 100 * c["coverage"]))
if "single" in result:
print("one message: median %.1f ms, p95 %.1f ms" % (result["single"]["median_ms"], result["single"]["p95_ms"]))
if __name__ == "__main__":
main()
|