dataforge/gliner_glirel_tuning
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore: initial sync — gliner+glirel benchmark notebooks

Browse Source 
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Egutierrez
2026-05-04 23:44:11 +02:00
commit b8c760d004
49 changed files with 47850 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
notebooks/01_gliner_glirel_tuning.ipynb
+865
View File
@@ -0,0 +1,865 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "6a7ef2a5",
"metadata": {},
"source": [
"# GLiNER + GLiREL — calibracion empirica\n",
"\n",
"**Objetivo:** entender empiricamente como funcionan **GLiNER** (entidades) y **GLiREL** (relaciones) para fijar thresholds operativos en el pipeline `extract_graph_hybrid` (panel _Paste & Extract_ de `graph_explorer`).\n",
"\n",
"**Hallazgo previo (sesion del merge 0013):** un solo `confidence_threshold=0.6` filtra GLiNER (0.92-0.99 facil) Y GLiREL (max 0.21 en el test). Resultado: el panel jamas muestra relaciones aunque GLiREL si las detecte. Este notebook valida la separacion necesaria de thresholds y mide rangos sanos.\n",
"\n",
"**Plan:**\n",
"1. Cargar modelos\n",
"2. **GLiNER** — barrido threshold sobre corpus EN/ES + sensibilidad a label sets\n",
"3. **GLiREL** — distribucion de scores sin filtro + sensibilidad a label phrasing\n",
"4. Recomendaciones operativas\n",
"\n",
"**Stack:** gliner==0.2.26, glirel==1.2.1, transformers==5.1, huggingface_hub==1.13. Modelos `urchade/gliner_multi-v2.1` (~600 MB) y `jackboyla/glirel-large-v0` (~1.5 GB), ambos cacheados en `~/.cache/huggingface/`."
]
},
{
"cell_type": "markdown",
"id": "2423c283",
"metadata": {},
"source": [
"## 1. Setup\n",
"\n",
"El kernel autocarga `FN_REGISTRY_ROOT` y anade `python/functions/` al `sys.path` (ver `.ipython/profile_default/startup/00_fn_registry.py`)."
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "67f48818",
"metadata": {
"execution": {
"iopub.execute_input": "2026-05-04T12:58:37.640753Z",
"iopub.status.busy": "2026-05-04T12:58:37.640602Z",
"iopub.status.idle": "2026-05-04T12:58:37.853224Z",
"shell.execute_reply": "2026-05-04T12:58:37.852377Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"FN_REGISTRY_ROOT: /home/lucas/fn_registry\n",
"results.json keys: ['gliner_threshold_sweep', 'glirel_score_distribution', 'glirel_topk_sweep', 'corpus', 'entity_labels', 'relation_labels']\n"
]
}
],
"source": [
"import os, sys, json, time, warnings\n",
"warnings.filterwarnings('ignore')\n",
"os.environ.setdefault('HF_HUB_DISABLE_PROGRESS_BARS', '1')\n",
"from pathlib import Path\n",
"\n",
"# Limpiar sys.path: el startup del kernel anade cada subdir de\n",
"# python/functions/ al top-level, y bigquery/datasets.py sombrea\n",
"# al paquete `datasets` de HuggingFace que necesita transformers.\n",
"# Dejamos solo el directorio padre 'python/functions/' para imports\n",
"# 'from datascience.gliner_load_model import ...' del estilo paquete.\n",
"_pf = '/home/lucas/fn_registry/python/functions'\n",
"sys.path = [p for p in sys.path if not (p.startswith(_pf + '/'))]\n",
"if _pf not in sys.path:\n",
" sys.path.insert(0, _pf)\n",
"\n",
"import pandas as pd\n",
"from datascience.gliner_load_model import gliner_load_model\n",
"from datascience.glirel_load_model import glirel_load_model\n",
"\n",
"RESULTS = json.loads(Path('../results.json').read_text())\n",
"print('FN_REGISTRY_ROOT:', os.environ.get('FN_REGISTRY_ROOT'))\n",
"print('results.json keys:', list(RESULTS.keys()))"
]
},
{
"cell_type": "markdown",
"id": "6dc6a22b",
"metadata": {},
"source": [
"## 2. Corpus de prueba\n",
"\n",
"4 textos cortos cubriendo dominios diferentes (ES/EN, corporativo/OSINT/journalism). Sirven para detectar drift de calidad por idioma y por tipo de contenido."
]
},
{
"cell_type": "markdown",
"id": "0f208d97",
"metadata": {},
"source": [
"### `es_corporate`\n",
"```\n",
"Pablo Isla, expresidente de Inditex, ha sido nombrado consejero de Telefonica. La operacion fue anunciada por el presidente Jose Maria Alvarez-Pallete en Madrid el pasado lunes. Inditex factura mas de 30.000 millones anuales y tiene su sede en Arteixo, A Coruna.\n",
"```\n",
"\n",
"### `en_corporate`\n",
"```\n",
"Pablo Isla, the former chairman of Inditex, has been appointed as a director of Telefonica. The announcement was made by Jose Maria Alvarez-Pallete, the chairman of Telefonica, in Madrid last Monday. Inditex has its headquarters in Arteixo, A Coruna.\n",
"```\n",
"\n",
"### `en_osint`\n",
"```\n",
"On 2024-08-15, attacker IP 185.220.101.45 connected to victim host 10.0.5.22 over TLS. Reverse DNS pointed to tor-exit-relay-3.onionrouter.net. Operator handle @phantomzero claimed responsibility on a forum. The C2 panel was hosted on hxxps://malwareops[.]biz/control behind Cloudflare.\n",
"```\n",
"\n",
"### `es_journalism`\n",
"```\n",
"Iberdrola y Endesa firmaron un acuerdo de colaboracion en proyectos eolicos en Galicia. El presidente de Iberdrola, Ignacio Galan, se reunio con la CEO de Endesa, Marina Serrano, en Bilbao. El acuerdo movilizara 2.000 millones de euros en cinco anos.\n",
"```\n"
]
},
{
"cell_type": "markdown",
"id": "8cbf0f22",
"metadata": {},
"source": [
"## 3. Carga de modelos\n",
"\n",
"Cold load: ~50s por modelo (descarga). Warm: ~8s. Cache global por (model_name, device)."
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "cf04dfad",
"metadata": {
"execution": {
"iopub.execute_input": "2026-05-04T12:58:37.855378Z",
"iopub.status.busy": "2026-05-04T12:58:37.855198Z",
"iopub.status.idle": "2026-05-04T12:58:52.254428Z",
"shell.execute_reply": "2026-05-04T12:58:52.253490Z"
}
},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"\u001b[0;93m2026-05-04 14:58:38.910665577 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under \"/sys/class/drm/card0\": device_discovery.cc:93 ReadFileContents Failed to open file: \"/sys/class/drm/card0/device/vendor\"\u001b[m\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"\u001b[1mDebertaV2Model LOAD REPORT\u001b[0m from: microsoft/deberta-v3-large\n",
"Key | Status | | \n",
"----------------------------------------+------------+--+-\n",
"mask_predictions.LayerNorm.bias | UNEXPECTED | | \n",
"lm_predictions.lm_head.bias | UNEXPECTED | | \n",
"lm_predictions.lm_head.LayerNorm.weight | UNEXPECTED | | \n",
"lm_predictions.lm_head.dense.weight | UNEXPECTED | | \n",
"lm_predictions.lm_head.dense.bias | UNEXPECTED | | \n",
"mask_predictions.classifier.bias | UNEXPECTED | | \n",
"mask_predictions.dense.weight | UNEXPECTED | | \n",
"mask_predictions.LayerNorm.weight | UNEXPECTED | | \n",
"mask_predictions.dense.bias | UNEXPECTED | | \n",
"mask_predictions.classifier.weight | UNEXPECTED | | \n",
"lm_predictions.lm_head.LayerNorm.bias | UNEXPECTED | | \n",
"\n",
"\u001b[3mNotes:\n",
"- UNEXPECTED\u001b[3m\t:can be ignored when loading from different task/architecture; not ok if you expect identical arch.\u001b[0m\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"GLiNER ready in 8.2s\n",
"GLiREL ready in 6.2s\n"
]
}
],
"source": [
"t0 = time.time(); gliner = gliner_load_model(); t_gliner = time.time()-t0\n",
"t0 = time.time(); glirel = glirel_load_model(); t_glirel = time.time()-t0\n",
"print(f'GLiNER ready in {t_gliner:.1f}s')\n",
"print(f'GLiREL ready in {t_glirel:.1f}s')"
]
},
{
"cell_type": "markdown",
"id": "08107c78",
"metadata": {},
"source": [
"## 4. GLiNER — barrido de threshold\n",
"\n",
"Para cada (corpus, label_set) corremos `predict_entities(threshold=0.0)` y filtramos a posteriori a {0.1, 0.3, 0.5, 0.7, 0.9}. Asi vemos la distribucion completa de scores sin recargar modelo."
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "46598320",
"metadata": {
"execution": {
"iopub.execute_input": "2026-05-04T12:58:52.257688Z",
"iopub.status.busy": "2026-05-04T12:58:52.257083Z",
"iopub.status.idle": "2026-05-04T12:58:52.284240Z",
"shell.execute_reply": "2026-05-04T12:58:52.283211Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"<div>\n",
"<style scoped>\n",
" .dataframe tbody tr th:only-of-type {\n",
" vertical-align: middle;\n",
" }\n",
"\n",
" .dataframe tbody tr th {\n",
" vertical-align: top;\n",
" }\n",
"\n",
" .dataframe thead th {\n",
" text-align: right;\n",
" }\n",
"</style>\n",
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: right;\">\n",
" <th></th>\n",
" <th>corpus</th>\n",
" <th>labels</th>\n",
" <th>t=.1</th>\n",
" <th>t=.3</th>\n",
" <th>t=.5</th>\n",
" <th>t=.7</th>\n",
" <th>t=.9</th>\n",
" <th>max_score</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th>0</th>\n",
" <td>es_corporate</td>\n",
" <td>generic_en</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>0.994</td>\n",
" </tr>\n",
" <tr>\n",
" <th>1</th>\n",
" <td>es_corporate</td>\n",
" <td>generic_es</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>0.990</td>\n",
" </tr>\n",
" <tr>\n",
" <th>2</th>\n",
" <td>en_corporate</td>\n",
" <td>generic_en</td>\n",
" <td>9</td>\n",
" <td>9</td>\n",
" <td>9</td>\n",
" <td>9</td>\n",
" <td>9</td>\n",
" <td>0.995</td>\n",
" </tr>\n",
" <tr>\n",
" <th>3</th>\n",
" <td>en_corporate</td>\n",
" <td>specific_en</td>\n",
" <td>9</td>\n",
" <td>9</td>\n",
" <td>9</td>\n",
" <td>9</td>\n",
" <td>8</td>\n",
" <td>0.991</td>\n",
" </tr>\n",
" <tr>\n",
" <th>4</th>\n",
" <td>en_osint</td>\n",
" <td>generic_en</td>\n",
" <td>12</td>\n",
" <td>6</td>\n",
" <td>1</td>\n",
" <td>0</td>\n",
" <td>0</td>\n",
" <td>0.604</td>\n",
" </tr>\n",
" <tr>\n",
" <th>5</th>\n",
" <td>en_osint</td>\n",
" <td>osint_en</td>\n",
" <td>13</td>\n",
" <td>8</td>\n",
" <td>6</td>\n",
" <td>2</td>\n",
" <td>2</td>\n",
" <td>0.953</td>\n",
" </tr>\n",
" <tr>\n",
" <th>6</th>\n",
" <td>es_journalism</td>\n",
" <td>generic_en</td>\n",
" <td>9</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>0.995</td>\n",
" </tr>\n",
" <tr>\n",
" <th>7</th>\n",
" <td>es_journalism</td>\n",
" <td>generic_es</td>\n",
" <td>9</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>8</td>\n",
" <td>7</td>\n",
" <td>0.992</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n",
"</div>"
],
"text/plain": [
" corpus labels t=.1 t=.3 t=.5 t=.7 t=.9 max_score\n",
"0 es_corporate generic_en 8 8 8 8 8 0.994\n",
"1 es_corporate generic_es 8 8 8 8 8 0.990\n",
"2 en_corporate generic_en 9 9 9 9 9 0.995\n",
"3 en_corporate specific_en 9 9 9 9 8 0.991\n",
"4 en_osint generic_en 12 6 1 0 0 0.604\n",
"5 en_osint osint_en 13 8 6 2 2 0.953\n",
"6 es_journalism generic_en 9 8 8 8 8 0.995\n",
"7 es_journalism generic_es 9 8 8 8 7 0.992"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"from datascience.gliner_load_model import gliner_load_model\n",
"thresholds = [0.1, 0.3, 0.5, 0.7, 0.9]\n",
"rows = []\n",
"for corpus_key, cdata in RESULTS['gliner_threshold_sweep'].items():\n",
" for ls_key, sdata in cdata.items():\n",
" scored = sdata['scored_at_t0']\n",
" max_s = max((s[2] for s in scored), default=0.0)\n",
" rows.append([corpus_key, ls_key, *[len(sdata[f't={t}']) for t in thresholds], round(max_s,3)])\n",
"df = pd.DataFrame(rows, columns=['corpus','labels','t=.1','t=.3','t=.5','t=.7','t=.9','max_score'])\n",
"df"
]
},
{
"cell_type": "markdown",
"id": "eed12fb4",
"metadata": {},
"source": [
"**Lectura:**\n",
"\n",
"- En **narrativa estructurada** (corporate, journalism), GLiNER da 8-9 entidades estables con scores 0.92-0.99. **`threshold=0.5` o `0.7` son seguros**, casi no se mueve el conteo.\n",
"- En **OSINT** (IPs, dominios, URLs) con labels genericas (`person`, `organization`...): scores _se hunden_ a max 0.60. **Cae todo a t=0.5**.\n",
"- Mismo OSINT con labels especificas (`ip_address`, `domain`, `url`): max 0.95, threshold 0.5 retiene 6.\n",
"- ES vs EN: practicamente identicos. El `gliner_multi-v2.1` es genuinamente multilingue. **Las labels EN funcionan igual de bien sobre texto ES.**\n",
"\n",
"**Conclusion 1:** `entity_threshold = 0.5` es seguro como default. Pero el **label set debe encajar al dominio** — una mala eleccion mata mas que un threshold mal puesto."
]
},
{
"cell_type": "markdown",
"id": "fed8f100",
"metadata": {},
"source": [
"### 4.1 Entidades concretas (en_corporate, generic_en, t=0.5)\n",
"\n",
"Para verificar que no son ruido."
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "5358e303",
"metadata": {
"execution": {
"iopub.execute_input": "2026-05-04T12:58:52.286116Z",
"iopub.status.busy": "2026-05-04T12:58:52.285916Z",
"iopub.status.idle": "2026-05-04T12:58:52.300382Z",
"shell.execute_reply": "2026-05-04T12:58:52.299264Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"<div>\n",
"<style scoped>\n",
" .dataframe tbody tr th:only-of-type {\n",
" vertical-align: middle;\n",
" }\n",
"\n",
" .dataframe tbody tr th {\n",
" vertical-align: top;\n",
" }\n",
"\n",
" .dataframe thead th {\n",
" text-align: right;\n",
" }\n",
"</style>\n",
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: right;\">\n",
" <th></th>\n",
" <th>text</th>\n",
" <th>label</th>\n",
" <th>score</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th>0</th>\n",
" <td>Pablo Isla</td>\n",
" <td>person</td>\n",
" <td>0.989302</td>\n",
" </tr>\n",
" <tr>\n",
" <th>1</th>\n",
" <td>Inditex</td>\n",
" <td>organization</td>\n",
" <td>0.992379</td>\n",
" </tr>\n",
" <tr>\n",
" <th>2</th>\n",
" <td>Telefonica</td>\n",
" <td>organization</td>\n",
" <td>0.992698</td>\n",
" </tr>\n",
" <tr>\n",
" <th>3</th>\n",
" <td>Jose Maria Alvarez-Pallete</td>\n",
" <td>person</td>\n",
" <td>0.975533</td>\n",
" </tr>\n",
" <tr>\n",
" <th>4</th>\n",
" <td>Telefonica</td>\n",
" <td>organization</td>\n",
" <td>0.990853</td>\n",
" </tr>\n",
" <tr>\n",
" <th>5</th>\n",
" <td>Madrid</td>\n",
" <td>location</td>\n",
" <td>0.966069</td>\n",
" </tr>\n",
" <tr>\n",
" <th>6</th>\n",
" <td>Inditex</td>\n",
" <td>organization</td>\n",
" <td>0.994649</td>\n",
" </tr>\n",
" <tr>\n",
" <th>7</th>\n",
" <td>Arteixo</td>\n",
" <td>location</td>\n",
" <td>0.968921</td>\n",
" </tr>\n",
" <tr>\n",
" <th>8</th>\n",
" <td>A Coruna</td>\n",
" <td>location</td>\n",
" <td>0.920429</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n",
"</div>"
],
"text/plain": [
" text label score\n",
"0 Pablo Isla person 0.989302\n",
"1 Inditex organization 0.992379\n",
"2 Telefonica organization 0.992698\n",
"3 Jose Maria Alvarez-Pallete person 0.975533\n",
"4 Telefonica organization 0.990853\n",
"5 Madrid location 0.966069\n",
"6 Inditex organization 0.994649\n",
"7 Arteixo location 0.968921\n",
"8 A Coruna location 0.920429"
]
},
"execution_count": 4,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"ents = RESULTS['gliner_threshold_sweep']['en_corporate']['generic_en']['t=0.5']\n",
"pd.DataFrame(ents, columns=['text','label','score','start','end'])[['text','label','score']]"
]
},
{
"cell_type": "markdown",
"id": "f4019283",
"metadata": {},
"source": [
"## 5. GLiREL — distribucion de scores\n",
"\n",
"Aqui esta el quid del bug: pasamos `threshold=0.0`, `top_k=5` y vemos los scores naturales que emite GLiREL. Comparamos dos estilos de label:\n",
"\n",
"- `snake_short`: `works_at`, `located_in`, `appointed_as`, ...\n",
"- `natural_long`: `person works at organization`, ...\n",
"\n",
"El folklore dice que el segundo deberia funcionar mejor (porque GLiREL es tipo zero-shot). Vamos a ver."
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "b0516987",
"metadata": {
"execution": {
"iopub.execute_input": "2026-05-04T12:58:52.302264Z",
"iopub.status.busy": "2026-05-04T12:58:52.302062Z",
"iopub.status.idle": "2026-05-04T12:58:52.313997Z",
"shell.execute_reply": "2026-05-04T12:58:52.312964Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"<div>\n",
"<style scoped>\n",
" .dataframe tbody tr th:only-of-type {\n",
" vertical-align: middle;\n",
" }\n",
"\n",
" .dataframe tbody tr th {\n",
" vertical-align: top;\n",
" }\n",
"\n",
" .dataframe thead th {\n",
" text-align: right;\n",
" }\n",
"</style>\n",
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: right;\">\n",
" <th></th>\n",
" <th>corpus</th>\n",
" <th>n_ents</th>\n",
" <th>label_style</th>\n",
" <th>n_rels</th>\n",
" <th>max_score</th>\n",
" <th>median_score</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th>0</th>\n",
" <td>es_corporate</td>\n",
" <td>8</td>\n",
" <td>snake_short</td>\n",
" <td>280</td>\n",
" <td>0.169</td>\n",
" <td>0.017</td>\n",
" </tr>\n",
" <tr>\n",
" <th>1</th>\n",
" <td>es_corporate</td>\n",
" <td>8</td>\n",
" <td>natural_long</td>\n",
" <td>280</td>\n",
" <td>0.061</td>\n",
" <td>0.010</td>\n",
" </tr>\n",
" <tr>\n",
" <th>2</th>\n",
" <td>en_corporate</td>\n",
" <td>9</td>\n",
" <td>snake_short</td>\n",
" <td>360</td>\n",
" <td>0.233</td>\n",
" <td>0.016</td>\n",
" </tr>\n",
" <tr>\n",
" <th>3</th>\n",
" <td>en_corporate</td>\n",
" <td>9</td>\n",
" <td>natural_long</td>\n",
" <td>360</td>\n",
" <td>0.080</td>\n",
" <td>0.007</td>\n",
" </tr>\n",
" <tr>\n",
" <th>4</th>\n",
" <td>es_journalism</td>\n",
" <td>8</td>\n",
" <td>snake_short</td>\n",
" <td>280</td>\n",
" <td>0.195</td>\n",
" <td>0.011</td>\n",
" </tr>\n",
" <tr>\n",
" <th>5</th>\n",
" <td>es_journalism</td>\n",
" <td>8</td>\n",
" <td>natural_long</td>\n",
" <td>280</td>\n",
" <td>0.138</td>\n",
" <td>0.007</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n",
"</div>"
],
"text/plain": [
" corpus n_ents label_style n_rels max_score median_score\n",
"0 es_corporate 8 snake_short 280 0.169 0.017\n",
"1 es_corporate 8 natural_long 280 0.061 0.010\n",
"2 en_corporate 9 snake_short 360 0.233 0.016\n",
"3 en_corporate 9 natural_long 360 0.080 0.007\n",
"4 es_journalism 8 snake_short 280 0.195 0.011\n",
"5 es_journalism 8 natural_long 280 0.138 0.007"
]
},
"execution_count": 5,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"rows=[]\n",
"for corpus, cdata in RESULTS['glirel_score_distribution'].items():\n",
" n_ents = len(cdata.get('entities', []))\n",
" for style, rels in cdata.get('styles', {}).items():\n",
" if isinstance(rels, list) and rels:\n",
" scores = sorted([r['score'] for r in rels], reverse=True)\n",
" rows.append([corpus, n_ents, style, len(rels), round(scores[0],3), round(scores[len(scores)//2],3)])\n",
" else:\n",
" rows.append([corpus, n_ents, style, 0, 0.0, 0.0])\n",
"df = pd.DataFrame(rows, columns=['corpus','n_ents','label_style','n_rels','max_score','median_score'])\n",
"df"
]
},
{
"cell_type": "markdown",
"id": "80cb8f95",
"metadata": {},
"source": [
"**Lectura — dos sorpresas:**\n",
"\n",
"1. **`snake_short` >> `natural_long`** por un factor 3-4×. Pasar `\"person works at organization\"` baja el score max de 0.23 a 0.08. **GLiREL fue entrenado con etiquetas estilo Wikipedia** (`P54`, `member_of_political_party`...), no con frases naturales. El prompt-engineering aqui es _menos_ es _mas_.\n",
"2. **EN > ES por ~25%**: `en_corporate` max 0.233 vs `es_corporate` max 0.169 con el mismo contenido factico. GLiREL tiene mejor cobertura del ingles.\n",
"3. **Texto OSINT** dio 0 entidades en GLiNER multi-v2.1 con labels genericas → no hay pares para GLiREL. (Para OSINT habria que cambiar GLiNER -> regex (que ya cubre IoCs) y dejar GLiREL para narrativa).\n",
"\n",
"**Conclusion 2:** **`relation_threshold` debe estar en 0.10-0.15**, NO en 0.6. El `confidence_threshold` global del pipeline debe partirse en dos."
]
},
{
"cell_type": "markdown",
"id": "e535e84b",
"metadata": {},
"source": [
"### 5.1 Efecto de `top_k`\n",
"\n",
"Subir `top_k` ¿descubre relaciones nuevas o solo añade ruido?"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "cc6855a0",
"metadata": {
"execution": {
"iopub.execute_input": "2026-05-04T12:58:52.315945Z",
"iopub.status.busy": "2026-05-04T12:58:52.315750Z",
"iopub.status.idle": "2026-05-04T12:58:52.325915Z",
"shell.execute_reply": "2026-05-04T12:58:52.324821Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"<div>\n",
"<style scoped>\n",
" .dataframe tbody tr th:only-of-type {\n",
" vertical-align: middle;\n",
" }\n",
"\n",
" .dataframe tbody tr th {\n",
" vertical-align: top;\n",
" }\n",
"\n",
" .dataframe thead th {\n",
" text-align: right;\n",
" }\n",
"</style>\n",
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: right;\">\n",
" <th></th>\n",
" <th>top_k</th>\n",
" <th>n_total</th>\n",
" <th>max</th>\n",
" <th>median</th>\n",
" <th>min</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th>0</th>\n",
" <td>top_k=1</td>\n",
" <td>72</td>\n",
" <td>0.233</td>\n",
" <td>0.129</td>\n",
" <td>0.036</td>\n",
" </tr>\n",
" <tr>\n",
" <th>1</th>\n",
" <td>top_k=3</td>\n",
" <td>216</td>\n",
" <td>0.233</td>\n",
" <td>0.045</td>\n",
" <td>0.003</td>\n",
" </tr>\n",
" <tr>\n",
" <th>2</th>\n",
" <td>top_k=5</td>\n",
" <td>360</td>\n",
" <td>0.233</td>\n",
" <td>0.016</td>\n",
" <td>0.000</td>\n",
" </tr>\n",
" <tr>\n",
" <th>3</th>\n",
" <td>top_k=10</td>\n",
" <td>360</td>\n",
" <td>0.233</td>\n",
" <td>0.016</td>\n",
" <td>0.000</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n",
"</div>"
],
"text/plain": [
" top_k n_total max median min\n",
"0 top_k=1 72 0.233 0.129 0.036\n",
"1 top_k=3 216 0.233 0.045 0.003\n",
"2 top_k=5 360 0.233 0.016 0.000\n",
"3 top_k=10 360 0.233 0.016 0.000"
]
},
"execution_count": 6,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"rows=[]\n",
"for tk, rels in RESULTS['glirel_topk_sweep']['by_topk'].items():\n",
" s = sorted([r['score'] for r in rels], reverse=True)\n",
" rows.append([tk, len(rels), round(s[0],3), round(s[len(s)//2],3), round(s[-1],3)])\n",
"df = pd.DataFrame(rows, columns=['top_k','n_total','max','median','min'])\n",
"df"
]
},
{
"cell_type": "markdown",
"id": "52f63ef3",
"metadata": {},
"source": [
"**Lectura:** `max` no se mueve. Solo crece `n_total` con peor score. **`top_k=1` o `top_k=3` es suficiente** para la app — subirlo solo añade ruido por debajo del threshold.\n",
"\n",
"**Conclusion 3:** dejar `top_k=1` por defecto en el panel. Si el usuario quiere ver alternativas, abrir un control avanzado."
]
},
{
"cell_type": "markdown",
"id": "163a20d2",
"metadata": {},
"source": [
"## 6. Recomendaciones operativas\n",
"\n",
"### Para `extract_graph_hybrid` y `paste_extract`\n",
"\n",
"| Param | Valor recomendado | Razon |\n",
"|---|---|---|\n",
"| `entity_threshold` | **0.50** (general) / **0.70** (narrativa estructurada) | GLiNER da 0.92-0.99 en narrativa; 0.5 deja margen para casos limite |\n",
"| `relation_threshold` | **0.15** (EN) / **0.10** (ES) | GLiREL tiene scores naturalmente bajos; 0.6 es absurdo |\n",
"| `top_k` | **1** | Subirlo solo añade peor evidencia |\n",
"| `relation_labels` | **snake_case corto** (`works_at`) | Frases naturales empeoran scores 3-4× |\n",
"| `entity_labels` | **dominio-especificas si OSINT** | Labels genericas hunden recall en texto OSINT |\n",
"\n",
"### Cambios concretos en el codigo\n",
"\n",
"1. **Issue nuevo en `graph_explorer`** — `0041-split-confidence-thresholds.md`:\n",
" - En `python/functions/pipelines/extract_graph_hybrid.py`: separar `confidence_threshold` en `entity_threshold` y `relation_threshold`.\n",
" - En `enrichers/paste_extract/run.py`: aceptar ambos parametros desde el manifest/ctx.\n",
" - En el panel C++ (`extract_panel.cpp`): dos sliders en lugar de uno, defaults 0.50 y 0.15.\n",
"2. **Test pytest existente** (`tests/test_paste_extract.py`) ya monkeypatchea el pipeline; añadir un test del path real con threshold separado cuando los modelos esten disponibles (skip si no).\n",
"3. **Documentar en `app.md`** que el path hybrid descarga ~2 GB la primera vez y queda en `~/.cache/huggingface/`.\n",
"\n",
"### Decisiones que NO se confirman aqui\n",
"\n",
"- Que pasa con texto > 512 tokens (GLiNER tiene window). Ver `extract_graph_hybrid` que ya hace chunking.\n",
"- Calidad real con LLM fallback activo (no probado en este notebook).\n",
"- Comportamiento con corpus mucho mas grande (este analysis prueba 4 textos cortos)."
]
},
{
"cell_type": "markdown",
"id": "1546f0f8",
"metadata": {},
"source": [
"## 7. Apendice — script reproducible\n",
"\n",
"Los datos vienen de `../results.json`, generado por `../run_experiments.py`. Para regenerar (cambiar corpus, labels, etc.):\n",
"\n",
"```bash\n",
"cd analysis/gliner_glirel_tuning\n",
"./.venv/bin/python3 run_experiments.py # ~30s con modelos calientes\n",
"./.venv/bin/python3 build_notebook.py # rebuild .ipynb con outputs\n",
"```"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.13.7"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
notebooks/02_e2e_spanish_graph.ipynb
+729
View File
File diff suppressed because one or more lines are too long
notebooks/03_mrebel_vs_glirel.ipynb
+794
View File
File diff suppressed because one or more lines are too long
notebooks/04_gliner2_winner.ipynb
+695
View File
File diff suppressed because one or more lines are too long
notebooks/05_long_text_and_pdf.ipynb
+1763
View File
File diff suppressed because one or more lines are too long
notebooks/06_improvements.ipynb
+419
View File
@@ -0,0 +1,419 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "4a6738d5",
"metadata": {},
"source": [
"# Mejoras al pipeline GLiNER2 sobre PDF — resultados empiricos\n",
"\n",
"**Pregunta:** del notebook 05 nos quedamos con un grafo de PDF con 382 entidades pero solo 48 aristas y 324 nodos aislados. **¿Como subimos las relaciones correctas y reducimos aislados?**\n",
"\n",
"Tras leer la API real de GLiNER2 (no la del README), identifique 6 palancas:\n",
"\n",
"1. `threshold` (default 0.5) — bajar a 0.3 / 0.2\n",
"2. `relations({type: description})` — pasar dict con descripciones, no lista\n",
"3. `batch_extract` con `batch_size=8`\n",
"4. Coreference simple (normalizacion + substring) entre chunks\n",
"5. Sliding window de 2 frases entre chunks\n",
"6. Limpieza del PDF (page numbers, saltos espurios)\n",
"\n",
"Ejecutado el benchmark en `run_improvements.py` y guardado en `improvements.json`. Este notebook solo carga los datos y los presenta — sin recargar GLiNER2."
]
},
{
"cell_type": "markdown",
"id": "ebbdc3f9",
"metadata": {},
"source": [
"## 0. Setup"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "c0adf6b4",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"keys: ['meta', 'configs', 'coref', 'top_entities_post_coref', 'top_relations_post_coref', 'ents_merged', 'rels_merged']\n"
]
}
],
"source": [
"import json\n",
"from pathlib import Path\n",
"import pandas as pd\n",
"DATA = json.loads(Path('../improvements.json').read_text())\n",
"print('keys:', list(DATA.keys()))"
]
},
{
"cell_type": "markdown",
"id": "59413647",
"metadata": {},
"source": [
"## 1. Pre-procesado del PDF (mejoras #5 y #6)\n",
"\n",
"Limpieza (`1/20` headers, saltos en medio de palabras, espacios duplicados) + chunking con sliding window de 2 frases."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "54e98462",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"raw chars: 89,882\n",
"clean chars: 88,714\n",
"chunks (overlap=2): 97\n",
"chunks (overlap=0): 66\n",
"\n",
"--- primeras 600 chars del clean ---\n",
"Banco Bilbao Vizcaya Argentaria, S.A., con domicilio en la Plaza San Nicolás, número 4, 48005 Bilbao,inscrito en el Registro Mercantil de Vizcaya, al tomo 2.083, Folio 1, Hoja BI-17-A, Inscripción 1ª con C.I.F. A-48265169POLÍTICA DE PROTECCIÓN DE DATOS PERSONALES 1. Política de Protección de Datos Personales T ómate tu tiempo y lee atentamente este documento. No dudes en pedirnos aclaraciones de lo que no entiendas.\n",
"En este apartado te explicamos para qué utilizará BBVA tus datos y, entre otros aspectos, qué derechos tienes relacionados con su uso.\n",
"INFORMACIÓN BÁSICA SOBRE PROTECCIÓN DE DATOS \n"
]
}
],
"source": [
"meta = DATA['meta']\n",
"print(f\"raw chars: {meta['raw_chars']:,}\")\n",
"print(f\"clean chars: {meta['clean_chars']:,}\")\n",
"print(f\"chunks (overlap=2): {meta['n_chunks_overlap']}\")\n",
"print(f\"chunks (overlap=0): {meta['n_chunks_no_overlap']}\")\n",
"print()\n",
"print('--- primeras 600 chars del clean ---')\n",
"print(meta['first_clean_600'])"
]
},
{
"cell_type": "markdown",
"id": "cfd5a2bd",
"metadata": {},
"source": [
"## 2. Bateria comparativa — 5 configuraciones\n",
"\n",
"Sobre los mismos 97 chunks del PDF cleaned + sliding window:\n",
"\n",
"| Config | threshold | schema | metodo |\n",
"|---|---|---|---|\n",
"| **A** baseline | 0.5 (default) | flat list | extract loop |\n",
"| **B** lower threshold | 0.3 | flat list | extract loop |\n",
"| **C** very low threshold | 0.2 | flat list | extract loop |\n",
"| **D** + descriptions | 0.3 | dict con desc | extract loop |\n",
"| **E** + batch | 0.3 | dict con desc | batch_extract |\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "4fecd7e7",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"config time ents rels edges isolates conn%\n",
"------------------- ------ ---- ---- ----- -------- -----\n",
"A: t=0.5 flat loop 134.3s 397 71 71 329 17.8%\n",
"B: t=0.3 flat loop 139.0s 517 204 204 389 26.0%\n",
"C: t=0.2 flat loop 133.9s 632 362 362 397 34.9%\n",
"D: t=0.3 desc loop 132.4s 517 204 204 389 26.0%\n",
"E: t=0.3 desc batch 163.6s 517 204 204 389 26.0%"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"rows = []\n",
"for c in DATA['configs']:\n",
" s = c['stats']\n",
" rows.append({\n",
" 'config': c['name'], 'time_s': c['elapsed'],\n",
" 'ents': s['n_ents'], 'rels': s['n_rels'], 'edges': s['n_edges'],\n",
" 'isolates': s['n_isolates'], 'conn_pct': s['connect_pct'],\n",
" })\n",
"df = pd.DataFrame(rows)\n",
"df"
]
},
{
"cell_type": "markdown",
"id": "757530b8",
"metadata": {},
"source": [
"**Lectura del benchmark:**\n",
"\n",
"- **Threshold es la palanca principal** y la unica que mueve la aguja:\n",
" - `0.5 → 0.3` = **+187% relaciones** (71 → 204)\n",
" - `0.3 → 0.2` = +78% mas (204 → 362), pero +22% entidades dudosas (517 → 632)\n",
" - **Sweet spot: 0.3** — gran ganancia sin meter ruido excesivo.\n",
"\n",
"- **Descripciones por relacion NO mejoran** este corpus legal denso (B = D, identico). Probable explicacion: GLiNER2 ya entiende los nombres cortos como `governed_by`, `subject_to` directamente. Las descripciones podrian pesar mas en relaciones ambiguas (`acquired` vs `merged_with`).\n",
"\n",
"- **batch_extract NO da speedup en CPU** — fue **25% mas lento** que el loop (E=163s vs D=132s). Sospecha: el modelo es CPU-bound y el batching introduce overhead sin paralelismo real (1 modelo, no caben 8 forward pass simultaneos en un core). Solo vale la pena con GPU.\n",
"\n",
"- **Sliding window de 2 frases** ya esta aplicado en TODOS los configs (forma parte del chunking). Su efecto exacto vs no-overlap requeriria una sexta config aparte (no medido aqui)."
]
},
{
"cell_type": "markdown",
"id": "98c616a6",
"metadata": {},
"source": [
"## 3. Coreferencia sobre la mejor config (E)\n",
"\n",
"Aplicamos un mergeo simple por:\n",
"\n",
"1. Lowercase + trim de puntuacion → cluster por nombre normalizado.\n",
"2. Substring match: nombres cortos absorbidos por largos del mismo tipo (`BBVA` ⊂ `Banco Bilbao Vizcaya Argentaria, S.A.`).\n",
"3. Re-escritura de relaciones para usar nombres canonicos.\n",
"\n",
"Coste: 0.62s. Tras coref:"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "def3dd7a",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"PRE-coref {'n_ents': 517, 'n_rels': 204, 'n_nodes': 526, 'n_edges': 204, 'n_isolates': 389, 'connected': 137, 'connect_pct': 26.0}\n",
"POST-coref {'n_ents': 401, 'n_rels': 166, 'n_nodes': 440, 'n_edges': 166, 'n_isolates': 318, 'connected': 122, 'connect_pct': 27.7}\n",
"absorbed: 72 aliases en 0.62s\n",
"\n",
"Samples de aliases absorbidos:\n",
" 'productos y servicios' → 'Información derivada de los productos y servicios contratados'\n",
" 'servicios contratados' → 'Información derivada de los productos y servicios contratados'\n",
" 'información' → 'Información derivada de los productos y servicios contratados'\n",
" 'productos' → 'Información derivada de los productos y servicios contratados'\n",
" 'servicios' → 'Información derivada de los productos y servicios contratados'\n",
" 'normativa' → 'normativa interna sobre prevención de crimen financiero'\n",
" 'blanqueo de capitales' → 'normativa de prevención del blanqueo de capitales'\n",
" 'interacción' → 'datos derivados de la interacción con chatbots'"
]
}
],
"source": [
"pre = DATA['coref']['pre_stats']\n",
"post = DATA['coref']['post_stats']\n",
"print('PRE-coref ', pre)\n",
"print('POST-coref', post)\n",
"print(f\"absorbed: {DATA['coref']['n_absorbed']} aliases en {DATA['coref']['elapsed']}s\")\n",
"print()\n",
"print('Samples de aliases absorbidos:')\n",
"for old, new in DATA['coref']['absorbed_sample']:\n",
" print(f' {old!r:55s} → {new!r}')"
]
},
{
"cell_type": "markdown",
"id": "5613c249",
"metadata": {},
"source": [
"**Lectura coref:**\n",
"\n",
"- **72 aliases absorbidos** en 0.62s — gratis para el usuario.\n",
"- Nodos: 526 → 440 (-86).\n",
"- Edges: 204 → 166 (-38) — _bajan porque las relaciones se mergean cuando ambos extremos colapsan al mismo canonico_.\n",
"- Aislados: 389 → 318 (-71, **-18%**).\n",
"- Conn%: 26.0% → 27.7% (mejora pequeña en porcentaje porque tambien se reducen los nodos totales).\n",
"\n",
"Lo que mas mejora la coreferencia es la **calidad del grafo**: en lugar de tener 5 nodos `productos`, `servicios`, `información`, etc. dispersos por el documento, los junta en una entidad canonica `Información derivada de los productos y servicios contratados`."
]
},
{
"cell_type": "markdown",
"id": "5d9af970",
"metadata": {},
"source": [
"## 4. Top entidades post-coref"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "fdb2f3c7",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"type canonical mentions n_aliases aliases_sample \n",
"------------- ------------------------------------------------------------ -------- --------- -----------------------------------------------------------------\n",
"organization BBVA Seguros 81 1 ['BBVA'] \n",
"data_category Datos Personales 47 0 [] \n",
"person cliente particular 34 1 ['cliente'] \n",
"organization Banco de España (CIRBE) 28 3 ['Banco de España', 'Banco', 'CIRBE'] \n",
"location Plaza San Nicolás 27 0 [] \n",
"location Vizcaya 22 0 [] \n",
"data_category datos derivados de la interacción con chatbots 19 3 ['interacción', 'chatbots', 'datos'] \n",
"law normativa interna sobre prevención de crimen financiero 19 1 ['normativa'] \n",
"right consentimiento 18 0 [] \n",
"data_category Datos transaccionales 18 1 ['transaccionales'] \n",
"data_category Información derivada de los productos y servicios contratado 17 5 ['productos y servicios', 'servicios contratados', 'información']\n",
"person clientes 15 0 [] \n",
"data_category Datos identificativos 14 0 [] \n",
"email derechosprotecciondatos@bbva.com 14 0 [] \n",
"data_category número de teléfono de contacto 13 1 ['contacto'] \n",
"person representante 12 0 [] \n",
"organization Agencia Española de Protección de Datos 12 0 [] \n",
"organization sociedades participadas 11 2 ['participadas', 'sociedades'] \n",
"person garante 11 0 [] \n",
"data_category Datos económicos 11 1 ['económicos'] "
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"rows = DATA['top_entities_post_coref'][:20]\n",
"df = pd.DataFrame(rows)\n",
"df"
]
},
{
"cell_type": "markdown",
"id": "36710c94",
"metadata": {},
"source": [
"## 5. Top relaciones post-coref"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "c5439813",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"from kind to count\n",
"---------------------------------------------- -------------- -------------------------------------------------- -----\n",
"BBVA Seguros governed_by Banco de España (CIRBE) 4 \n",
"Datos Personales protected_by Agencia Española de Protección de Datos 4 \n",
"Datos Personales protected_by Política de Protección de Datos Personales 3 \n",
"BBVA Seguros subject_to obligaciones legales 3 \n",
"derechos de acceso rights_against datos derivados de la interacción con chatbots 3 \n",
"contratación controlled_by BBVA Seguros 3 \n",
"BBVA Seguros subsidiary_of Grupo BBVA 2 \n",
"Datos Personales protected_by BBVA Seguros 2 \n",
"BBVA Seguros contact_for Información derivada de los productos y servicios 2 \n",
"Delegado de Protección de Datos contact_for BBVA Seguros 2 \n",
"BBVA Seguros controlled_by Banco de España (CIRBE) 2 \n",
"domicilio located_in Plaza San Nicolás 2 \n",
"datos de contacto contact_for clientes 2 \n",
"BBVA Seguros located_in España 2 \n",
"contratos de crédito inmobiliario governed_by Ley 5/2019 2 \n",
"Avda. de la Industria located_in MADRID 2 \n",
"bbva.es located_in MADRID 2 \n",
"datos derivados de la interacción con chatbots subject_to normativa interna sobre prevención de crimen finan 2 \n",
"Datos Personales subject_to normativa interna sobre prevención de crimen finan 2 \n",
"Emailage Corporation located_in Londres 2 "
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"rows = DATA['top_relations_post_coref'][:20]\n",
"df = pd.DataFrame(rows)\n",
"df"
]
},
{
"cell_type": "markdown",
"id": "3c830cb5",
"metadata": {},
"source": [
"## 6. Conclusion — recetario operativo\n",
"\n",
"**Para subir relaciones correctas y reducir aislados en GLiNER2 sobre PDF, en orden de impacto/coste:**\n",
"\n",
"| Mejora | Ganancia tipica | Coste de implementacion |\n",
"|---|---|---|\n",
"| ⭐ `threshold=0.3` (vs default 0.5) | **+187% relaciones** | 1 parametro |\n",
"| ⭐ Coreferencia simple (normalize + substring) | **-18% aislados** | ~30 lineas Python pure |\n",
"| Limpieza del PDF (`N/20`, saltos) | -1.3% chars de ruido + chunks mas estables | ~10 lineas regex |\n",
"| `threshold=0.2` (mas agresivo) | +78% relaciones extra, +22% ents dudosas | trade-off |\n",
"| ❌ Descripciones por relacion | Sin efecto en este corpus | dict en vez de list |\n",
"| ❌ batch_extract en CPU | 25% mas lento | API distinta |\n",
"| ❌ Sliding window con chunks de 1500 chars | Marginal | 5 lineas |\n",
"\n",
"**Stack final recomendado:**\n",
"\n",
"```python\n",
"# 1. Carga GLiNER2 (Apache 2.0)\n",
"model = GLiNER2.from_pretrained('fastino/gliner2-large-v1')\n",
"\n",
"# 2. Pre-procesa PDF\n",
"raw = extract_pdf_text(pdf_path) # registry: extract_pdf_text_py_core\n",
"clean = clean_pdf_text(raw) # NUEVA funcion del registry\n",
"chunks = chunk_with_overlap(clean, max_chars=1500, overlap_sentences=2) # NUEVA\n",
"\n",
"# 3. Schema + extract con threshold=0.3\n",
"schema = model.create_schema().entities([...]).relations([...])\n",
"results = [model.extract(c['text'], schema=schema, threshold=0.3) for c in chunks]\n",
"\n",
"# 4. Aggregate + coref\n",
"ents, rels = aggregate(results) # NUEVA, pura\n",
"ents, rels, _ = merge_aliases(ents, rels) # NUEVA, pura\n",
"```\n",
"\n",
"## Funciones a promover al registry (proximo fn-constructor)\n",
"\n",
"Aproximadamente **6 funciones nuevas**, casi todas puras:\n",
"\n",
"1. `gliner2_load_model_py_datascience` (impure) — Apache 2.0, NER+RE joint\n",
"2. `clean_pdf_text_py_core` (pure) — limpieza de artefactos PyPDF2\n",
"3. `chunk_with_overlap_py_core` (pure) — chunking con sliding window\n",
"4. `aggregate_extraction_results_py_core` (pure) — dedupe + counter\n",
"5. `merge_entity_aliases_py_core` (pure) — coref simple normalize + substring\n",
"6. `extract_graph_from_pdf_py_pipelines` (impure) — composicion completa\n",
"\n",
"Esto cierra el ciclo: el flujo del notebook se vuelve _una llamada del registry_ reusable cross-project."
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.13.7"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
notebooks/07_nuextract_vs_gliner2.ipynb
+933
View File
File diff suppressed because one or more lines are too long
notebooks/08_improving_gliner2.ipynb
+1419
View File
File diff suppressed because one or more lines are too long
notebooks/09_spacy_es_openie.ipynb
+1078
View File
File diff suppressed because one or more lines are too long