File size: 5,198 Bytes
4de12e0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
# Evaluating on HandleAtlas-benchmark

A self-contained eval harness for the
[`LumeData/HandleAtlas-benchmark`](https://huggingface.co/datasets/LumeData/HandleAtlas-benchmark)
dataset. The whole protocol — splits, labels, thresholds, scoring — lives in
[`eval.yaml`](./eval.yaml); the runner in [`run_eval.py`](./run_eval.py) reads
that file, loads the dataset from the Hub, and reports the same numbers used
on the model cards.

---

## What gets measured

| metric | definition |
|---|---|
| **span-only F1** *(primary)* | label-agnostic; a prediction matches a gold span when their character-offset IoU ≥ 0.5 |
| **span+label F1** | exact match on `(start, end, label)` — only fair when models share this taxonomy |
| **latency** | wall-clock ms per record, batch=1, CPU |

Per-label P/R/F1 is also printed for diagnostic use.

---

## Setup

You only need the deps for the adapter you plan to use.

```bash
# core (always needed)
pip install datasets pyyaml

# adapter: gliner  (GLiNER models, including HandleAtlas)
pip install gliner "transformers<5" torch
#   + ONNX (only if you pass --onnx):
pip install onnx onnxruntime optimum "accelerate>=0.26"

# adapter: hf-pipeline  (BERT/DeBERTa token-classification models)
pip install "transformers<5" torch
#   Some PII models (e.g. openai/privacy-filter) require a newer transformers:
pip install "git+https://github.com/huggingface/transformers.git"
```

Recommended: use `uv` so each run gets an isolated env.

```bash
uv run --with gliner --with 'transformers<5' --with torch \
       --with datasets --with pyyaml \
       python run_eval.py --adapter gliner --model LumeData/HandleAtlas-166m
```

---

## Quickstart

### Evaluate HandleAtlas (PyTorch float)

```bash
python run_eval.py --adapter gliner --model LumeData/HandleAtlas-166m
```

### Evaluate the CPU (INT8 ONNX) variant

```bash
python run_eval.py \
  --adapter gliner \
  --model LumeData/HandleAtlas-166m-CPU \
  --onnx
```

### Evaluate base GLiNER zero-shot

```bash
python run_eval.py --adapter gliner --model urchade/gliner_small-v2.1
```

### Evaluate a different-taxonomy PII model (label-agnostic only)

The span-only F1 is fair across taxonomies; the span+label F1 will read low —
that's expected.

```bash
python run_eval.py \
  --adapter hf-pipeline \
  --model openai/privacy-filter
```

If you want span+label to be meaningful, pass `--label-map` to translate the
model's classes into the HandleAtlas label set:

```bash
python run_eval.py \
  --adapter hf-pipeline \
  --model some/pii-model \
  --label-map "USERNAME=generic_username,SOCIAL=instagram_username"
```

---

## CLI flags

| flag | what it does |
|---|---|
| `--adapter` | `gliner` or `hf-pipeline` (required) |
| `--model` | HF Hub id or local path; passed straight to the adapter (required) |
| `--spec` | path to `eval.yaml` (defaults to the one next to `run_eval.py`) |
| `--threshold` | override the decoding threshold from the spec |
| `--label-map` | `model_label=target_label,...` for `hf-pipeline` |
| `--onnx`, `--onnx-file` | load the ONNX file instead of the PyTorch weights (GLiNER) |
| `--threads` | CPU threads (default 8) |
| `--device` | hf-pipeline: `-1`=CPU, `0`=first GPU |
| `--limit N` | only score the first N records (smoke test) |
| `--save-predictions PATH` | write per-record predictions to JSONL |
| `--json` | also dump the metrics JSON at the end |

---

## Adding your own model

If neither built-in adapter fits, subclass `Adapter` in `run_eval.py` and
register it. The contract is one method:

```python
class MyAdapter(Adapter):
    name = "mymodel"

    def __init__(self, model, labels, threshold, per_label_thresholds, **_):
        super().__init__(model, labels, threshold, per_label_thresholds)
        # load once
        self.m = my_lib.load(model)

    def predict(self, text: str) -> list[dict]:
        # MUST return [{"start": int, "end": int, "label": str}, ...]
        # `start`/`end` are Python character offsets, end exclusive.
        return [{"start": s, "end": e, "label": l}
                for s, e, l in self.m.extract(text, self.labels)]

ADAPTERS["mymodel"] = MyAdapter
```

Then:

```bash
python run_eval.py --adapter mymodel --model path/or/id
```

That's it — scoring and reporting are reused.

---

## Reproducing the published numbers

The values printed by this harness on `--model LumeData/HandleAtlas-166m`
should match the baselines table in `eval.yaml`. If they don't, check:

1. **Threshold** — the spec sets `0.5` with a `0.65` override on
   `generic_username`. Don't override unless you mean to.
2. **Split** — the harness only uses the `test` split; the published numbers
   are on this same split with `seed=123`.
3. **Hardware** — latency was measured on a MacBook Pro M5 Pro at 8 threads.
   Different hardware will give different ms numbers; F1 should not move.

---

## Files

| file | purpose |
|---|---|
| `eval.yaml` | declarative protocol — dataset, labels, thresholds, metrics, baselines |
| `run_eval.py` | runner — loads dataset, calls the adapter, scores, reports |
| `EVAL.md` | this file |
| `test.jsonl` | the actual eval data, also mirrored on the Hub |
| `README.md` | dataset card |