tagasenti_model

Model Description

tagasenti_model (v6 release) is a fine-tuned XLM-RoBERTa-large (355M parameters, 250K SentencePiece vocab) for ternary sentiment classification (Negative / Neutral / Positive) of Tagalog and Taglish (Tagalog-English code-switching) text.

It was trained on TagaSenti v6 — 35,686 rows (deduplicated to 34,945), the final dataset on Hub (jjjardev/tagasenti), and achieves 84.8% test accuracy and macro-F1 0.848 on the held-out Tagalog test set (stratified 80/10/10, seed 42, val peak 0.863 at step 2400, 3 epochs). Zero-shot transfer to Hiligaynon (HiliSenti) is 58.6% F1, with adversarial 100-set ceiling **83%**.

A quantized INT8 ONNX version (tagasenti_int8.onnx, ~537 MB, dynamic quantization per_channel=True, opset 18) is included in this repository for on-device / CPU-only deployment with no PyTorch required — see On-Device (ONNX).

Version note — why v6:

• Dataset on Hub is v6 — 35,686 rows (v5.1 34,287 + 1,399 selective: Contextual Pos 300, Neutral Factual 404, Hedged 300, Idioms 400). This release matches dataset versioning and fixes v4’s neutral hedging bias (“pero” → 0 err, Neu prediction error 0.9% vs 3 errors in v4).
• Project-best remains v4 — 32,179 rows (v3 28,766 + 3,413 “no-keywords”) with test F1 0.866, HiliSenti 0.624, val peak 0.878 at step 3200 (unsaved due to save_total_limit=2). v5.1 (34,287) is the best saved checkpoint (0.856, load_best_model_at_end success). Retraining on v6 yields 0.848 — the ceiling of template-based synthetic data (~0.86 / 83% / 0.624) is well-characterized in paper_notes/COMPLETE_DOCUMENTATION.md §6, §14–15.
• v6 is the final, dataset-aligned release — no further iterations planned. Use v4 weights from paper_notes/ if you need peak accuracy/cross-lingual; use v6 for reproducibility with the Hub dataset.

The model was developed solo in 9 days (June 10–21, 2026) on Colab free tier (Tesla T4) + LLM web UIs at $0 cost, using targeted adversarial augmentation over 4 generations to solve sarcasm, negation, and hedging failures.

Intended Uses

• Sentiment analysis of Tagalog/Taglish customer reviews (Shopee, Lazada, e-commerce), social media comments (TikTok-style), and news-adjacent text
• Social media monitoring for Filipino-speaking communities
• Trend analysis for product feedback, service reviews, and public opinion
• Research on low-resource Philippine languages, code-switching, and cross-lingual transfer (Tagalog → Hiligaynon/Cebuano)
• Educational and civic technology applications in the Philippines

Out-of-Scope Uses

• Fine-grained emotion detection (joy, anger, sarcasm) — model outputs only 3 coarse labels; adversarial sarcasm category targets sarcasm but does not label it
• Aspect-based sentiment analysis (ABSA)
• High-stakes moderation, legal, or financial decisions without human verification — idiom error rate is 47% (8/17 adversarial errors)
• Languages other than Tagalog/Taglish — zero-shot works partially for Hiligaynon (62.4%) but is not optimized for it; Cebuano/Ilokano not evaluated
• Long documents (>128 tokens) — truncated at 128 (p99=71–72); median HiliSenti is 96 tokens, Tagalog is ~63

Training Data

The model was trained on TagaSenti (jjjardev/tagasenti, DOI 10.57967/hf/9620, CC BY-SA 4.0). The dataset has two layers:

Layer 1 — Real-world base (25,000 rows, perfectly balanced 33/33/33, stratified 80/10/10 seed 42):

Source	Domain	Language	Rows	Label Type
FiReCS (ccosme/FiReCS) — Shopee/Lazada reviews	E-commerce	Tagalog/Taglish	10,481	3-class gold
SentiTaglish (ccosme/SentiTaglishProductsAndServices) — products & services (label 4→3 collapsed)	E-commerce	Tagalog/Taglish	6,816	4→3 remapped
Amazon Polarity (fancyzhx/amazon_polarity, 10k sample → 3,942 after 14–50w filter + 1k synthetic neutral via Gemini 3.1 Pro)	E-commerce (multi-category)	English→Tagalog (Gemini 3.1 Pro)	3,942	Binary→3 + synthetic
NewsPH (jcblaise/newsph, 43k→41,768→5,999 labeled via Deepseek V4 Pro, 20×300 batches)	News (politics 17%, weather 14%, economy 8%)	Tagalog (formal, 99% named entities)	5,999	Synthetic 3-class
TikTok Hate Speech (imperialite/filipino-tiktok-hatespeech, 18,538→17,260→11,961→6,292 via Deepseek label+rewrite, 10 batches)	Social media (TikTok comments, 54% Q/! )	Tagalog+English + slang	6,292	Synthetic 3-class (original hate labels discarded)

Layer 2 — Targeted adversarial augmentation (4 generations, 10,686 rows total; 9,378 after filtering):

Generation	Version	Rows	Categories	Strategy	Generator	Result
1st	v3	3,766	8 narrow (sarcasm, negation, wish, indirect complaint, factual-neutral trigger, mixed/ambiguous, rhetorical Qs, hard positive)	Keyword guardrails, batch rotation, human spot-checks	Gemini 3.1 Pro	+18pp benchmark (74%→92% 50-set)
2nd	v4	3,413	6 narrow (sarcasm 800, complaint 700, relief 400, double-neg 400, hedge 300, idiom 400)	“No keywords” — ban explicit sentiment words, force context-only learning	Gemini/Deepseek	Production — test +3.3 Hili, sarcasm 5→0
3rd	v5/v5.1	2,417	8 (11/28 files, single-batch → formulaic)	Idiom/pattern-heavy	Gemini/Deepseek	Failed to surpass v4 (0.856)
4th	v6	1,399	4 selective (Contextual Pos 300, Neutral Factual 404, Hedged 300, Idioms 400)	Class rebalancing (Pos/Neu/Idioms only)	Gemini/Deepseek	Neu bias fixed, 0.848

Final composition v6: 35,686 rows | Neg 13,259 (37.2%) / Neu 11,656 (32.7%) / Pos 10,771 (30.2%) | Class weights [0.897, 1.021, 1.104]. v4: 32,179 rows | Neg 12,067 / Neu 10,322 / Pos 9,790 | Weights [0.8889, 1.0392, 1.0956]. Deduplicated for training: ~34,945 (741 exact duplicates, 113 with conflicting labels — keep first, see scripts/validate_dataset.py).

Excluded: DailyDialog (CC BY-NC-SA 4.0 incompatible), iTANONG (DOST-ASTI gated), Shopee/SST-2/IMDb (domain saturation / phrase fragments / excessive length).

Training Procedure

• Base model: xlm-roberta-large (355M, 250k SentencePiece vocab) — cased, no lowercasing (preserves Filipino emphasis: AYAW KO NA ≠ ayaw ko na)
• Text normalization (normalize_tagalog in scripts/train/TagaSenti.py):
  • Unicode NFKC (ligatures/homoglyphs/half-width collapsed)
  • Laughter canonicalization: hahaha… / HAHAHA → hahaha (IGNORECASE), hehehe similarly
  • Reduplication: abot2 → abot abot (≥3-char words), abot-abot → abot abot (hyphenated)
  • Slang expansion — 25 high-confidence patterns (wla→wala, dko→di ko, kc/ksi→kasi, sya/xa/xia→siya, nya→niya, dba/diba→di ba, etc.); 3 removed for collision (pro→pero, mn→man, ala→wala)
  • Char deduplication ≥3→2 (graaaabe→graabe), whitespace collapse
  • No add_prefix_space (SentencePiece, not BPE)
• Tokenizer: AutoTokenizer.from_pretrained("xlm-roberta-large"), dynamic DataCollatorWithPadding, max length p99 = 71–72 (capped at 128) from 2,000-sample probe; HiliSenti p95=96
• Split: Stratified 80/10/10 by label, seed 42 (v4: 25,743/3,218/3,218; v6: ~27,955/3,569/3,569). Deduplicate before split to avoid leakage. Offline v1–v4 held adversarial train-only for comparable val/test; Hub single train split is mixed — scripts/train/TagaSenti.py does reproducible random split.
• Optimization:
  • Effective batch 32 (16 per device × 2 grad accum), LR 2e-5, cosine min_lr 1e-6, warmup 312 steps (10%), adamw_torch_fused, weight_decay 0.05
  • label_smoothing 0.10 via CustomTrainer (explicitly passed to nn.CrossEntropyLoss — fixes HF Trainer bug where label_smoothing_factor is ignored when compute_loss is overridden)
  • gradient_checkpointing True, fp16 when CUDA, 5 epochs v4 (4,025 steps) / 3 epochs v5.1/v6 (2,574/2,679 steps), early stopping patience 3 (f1_macro)
  • Class weights from compute_class_weight("balanced") per version; v4 [0.8889, 1.0392, 1.0956]
• Callbacks: EarlyStoppingCallback, ConfusionMatrixCallback (uses shared _last_eval cache from compute_metrics — fixes silent no-op where on_evaluate never receives eval_dataset)
• Infrastructure: Colab free tier Tesla T4, Drive 15GB limit → save_total_limit pruned best ckpt (v4 peak 0.878 not saved); v5.1 fixed with save_total_limit=0 (✅ load_best_model_at_end), v6 failed on sharded safetensors (save_safetensors=False needed). Tokenization cap 4 workers.
• Compute: ~50 min training + 30 min HiliSenti eval (v1 baseline); ~30h active work over 9 days total.

Evaluation Results

Tagalog Held-Out Test (v6 release, 3,569 rows, final step 2679 — val peak 2400 not loaded due to sharded safetensors)

Metric	Negative	Neutral	Positive	Overall
Precision	—	—	—	—
Recall	—	—	—	—
F1-Score	0.861	0.831	0.853	—
Accuracy	—	—	—	84.8%
Macro F1	—	—	—	0.848
Balanced Accuracy	—	—	—	84.8%

Per-class at peak step 2400 (F1 0.863, loss 0.555): Neg 0.882 / Neu 0.834 / Pos 0.871, pred 1,354/1,176/1,039 (true 1,190/1,190/1,189) — Neu error 0.9% (hedging bias fixed). Final step 2679 used for Hub weights: test 0.848 (Neg 0.861 / Neu 0.831 / Pos 0.853).

Val peak (v6, step 2400): F1 0.863 (Neg 0.882 / Neu 0.834 / Pos 0.871), loss 0.555. Project-best (v4, step 3200, unsaved): F1 0.878 (Neg 0.891 / Neu 0.848 / Pos 0.894).

Cross-version: v1 0.835 → v3 0.849 (+1.4) → v4 0.866 (+1.7, peak 0.878) → v5.1 0.856 (best saved, load_best_model_at_end ✅) → v6 0.848 (release, dataset-aligned). 83% adversarial ceiling holds.

Cross-Lingual Zero-Shot — HiliSenti (Hiligaynon, jjjardev/hilisenti-v1 test 2,242 rows, same normalize_tagalog)

Metric	v1	v3	v4 (peak)	v5.1	v6 (release)
F1 Macro	0.558	0.591	0.624	0.585	0.586
Neg F1	0.571	0.618	0.649	—	—
Neu F1	0.464	0.556	0.565	—	—
Pos F1	0.637	0.599	0.657	—	—
Accuracy	55.8%	59.2%	62.4%	—	58.6%

v4 Pos +5.8 is the single largest cross-lingual gain (context-only “relief” 400 rows transferred). v6 ties v5.1 on Hili (0.586); ~25% relative drop from Tagalog is the zero-shot ceiling; 500 Hiligaynon examples would likely push to 0.70–0.75.

Adversarial Benchmark (100 sentences, edge cases: idioms, double neg, sarcasm, hedging)

Model	Score	Sarcasm (5)	Idioms (10)	Neu Hedge	Negation
v3	81% (81/100)	5 err	10 err	1 err	2 err
v4	83% (83/100)	0 err	8 err	3 err	2 err
v5.1	83% (83/100)	0 err	8 err	0 err	4 err

50-sentence benchmark (v1 era): v1 74% → v2 68% (regression) → v3 92% (+18). 100-general (too easy) v3 97.9% superseded by adversarial set. v6 not independently benchmarked (~83%). Idioms are the #1 unsolved problem — 8 errors stable across v4/v5.1/v6 despite 3 generations of idiom data.

Version guidance (v6 is Hub release): Dataset-aligned / hedged neutral (customer feedback) → v6 (Neu 0.831, 0.9% error); peak accuracy / cross-lingual → v4 (0.866 / 0.624, unsaved peak 0.878) from paper_notes/; reproducibility / best saved → v5.1 (0.856). Idiom-heavy → none (8 errors stable).

Limitations

• Idioms: 8/17 adversarial errors (47%) are Filipino idioms/figurative — template-based generation cannot teach non-compositional semantics (ningas-kugon etc.). All 3 idiom generations failed identically.
• load_best_model_at_end unreliable: 3 of 4 runs failed (v4 save_total_limit=2 deleted peak 0.878; v5 10; v6 sharded safetensors). Only v5.1 succeeded (0 + pytorch_model.bin). Use save_safetensors=False in future.
• Single seed (42): No variance estimate; 3 seeds would confirm robustness (<0.01 std expected).
• LLM label noise: NewsPH (5,999) + TikTok (6,292) labeled by Deepseek V4 Pro single-pass, zero human validation; Amazon translated by Gemini 3.1 Pro with syntactic bottleneck (35% pero/kaso contrast). 500-sample human eval planned but never executed.
• Domain gaps: Movie reviews (SST-2/IMDb excluded), food/restaurant, beauty/cosmetics, education (Student 2,272 deferred) absent; 3 base domains only (e-commerce 63%, news 18%, social 19% at 4_READY).
• Translation artifacts: Amazon machine-translated Tagalog shows compressed syntax; adversarial data may contain non-native artefacts.
• No confidence calibration: Single label output, no distribution; neutral boundary weakest (lacks strong lexical signal).
• Truncation: 128-token cap may cut long news sentences (HiliSenti p95=96 vs Tagalog 63–72).
• Reproducibility: All LLM generation via manual web UI pastes (72 batches, ~27k rows) — free but not API-scripted; prompt templates in paper_notes/ + seed 42 mitigate.

Ethical Considerations

• Public data only: FiReCS/SentiTaglish (public reviews), Amazon (public), NewsPH (public news sites with named individuals/locations), TikTok (public comments) — all under open licenses or public availability. No private data.
• No PII scrubbing: Social media comments and news articles contain real names, locations, violence descriptions. HFREADME.md includes Important Notice — no anonymization applied. Users requiring anonymized data should apply their own NER/scrubbing pipeline.
• Adversarial sentences are synthetic — no personal information.
• Biases: Domain skew (e-commerce 35.6% at v6), LLM pretraining bias, translation bias, socioeconomic skew (TikTok/SentiTaglish demographics), adversarial generation bias (English-centric LLM). See dataset card Discussion of Biases.
• Intended effect: Lower barrier for Filipino NLP, spur Philippine language tech, demonstrate $0 solo researcher pipeline — not for high-stakes decisions without verification.
• Licensing is tri-partite: Dataset CC BY-SA 4.0 (ShareAlike required by TikTok), model weights Apache 2.0, code MIT — see LICENSES.md.

How to Use

Via Transformers

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

# Load model and tokenizer — downloads from https://huggingface.co/jjjardev/tagasenti_model
model = AutoModelForSequenceClassification.from_pretrained("jjjardev/tagasenti_model")
tokenizer = AutoTokenizer.from_pretrained("jjjardev/tagasenti_model")
model.eval()

# Example inference (Tagalog / Taglish)
sentences = [
    "Ang ganda ng quality ng tela, worth it ang price!",  # Positive
    "Wala pa ring update ang order ko hanggang ngayon.",  # Negative
    "Sa aking palagay, hindi naman ito gaanong importante.",  # Neutral
    "Ang bait mo naman, pinagbigyan mo ako sa wakas.",  # Negative (sarcasm — adversarial)
]
inputs = tokenizer(sentences, padding=True, truncation=True, max_length=128, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)
    preds = logits.argmax(dim=-1)

labels = ["Negative", "Neutral", "Positive"]
for sent, pred, prob in zip(sentences, preds, probs):
    print(f"{labels[pred]} ({prob[pred]:.2%}): {sent}")

Using the Pipeline API

from transformers import pipeline

classifier = pipeline("text-classification", model="jjjardev/tagasenti_model", tokenizer="jjjardev/tagasenti_model", truncation=True, max_length=128)

result = classifier("Napakabagal ng serbisyo nila, nakakainis.")
print(result)  # [{'label': 'Negative', 'score': 0.91}]

# Batch
results = classifier([
    "Sobrang saya ko sa nabili ko!",
    "Okay lang, parang kulang sa alat.",
    "Hindi ko nagustuhan, sira agad.",
])
for r in results:
    print(r)

On-Device (ONNX)

Download the INT8 ONNX model: https://huggingface.co/jjjardev/tagasenti_model/resolve/main/tagasenti_int8.onnx

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("jjjardev/tagasenti_model")
session = ort.InferenceSession("tagasenti_int8.onnx", providers=["CPUExecutionProvider"])

labels = ["Negative", "Neutral", "Positive"]
text = "Ang ganda ng quality ng tela, worth it ang price!"
inputs = tokenizer(text, return_tensors="np", truncation=True, max_length=128)
logits = session.run(["logits"], {
    "input_ids": inputs["input_ids"].astype(np.int64),
    "attention_mask": inputs["attention_mask"].astype(np.int64),
})[0]
print(labels[logits.argmax().item()])  # Positive

ONNX spec: inputs input_ids (int64) and attention_mask (int64) with dynamic [batch, seq_len]; output logits (float32 [batch, 3]). Exported from the v6 weights via FP32 ONNX (opset 18, dynamic batch/sequence) + INT8 dynamic quantization (QuantType.QInt8, per_channel=True, single file, no external .data). Spot-checked against the FP32 model on 4 Tagalog sentences (Positive / Negative / Neutral + sarcasm) — all 4 labels match. Model weights remain Apache 2.0.

CLI (local repo)

pip install -r requirements.txt  # or pip install -e .
python scripts/inference.py "Ang ganda ng tela!" --hf-id jjjardev/tagasenti_model
python scripts/inference.py --file sentences.txt --json --batch-size 16
python scripts/validate_dataset.py --csv tagasenti_dataset.csv

Citation

If you use this model or TagaSenti in your research, please cite:

Model:

@misc{jessie_james_jarder_2026_tagasenti_model,
  author    = {Jessie James Jarder},
  title     = {tagasenti_model (Revision tagasenti_model)},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/jjjardev/tagasenti_model},
  doi       = {10.57967/hf/9620},
}

Dataset:

@misc{jessie_james_jarder_2026,
  author       = {Jessie James Jarder},
  title        = {tagasenti (Revision 3ebda33)},
  year         = {2026},
  url          = {https://huggingface.co/datasets/jjjardev/tagasenti},
  doi          = {10.57967/hf/9620},
  publisher    = {Hugging Face}
}

APA (dataset):

Jarder, J. J. (2026). tagasenti (Revision 3ebda33) [Data set]. Hugging Face. https://doi.org/10.57967/hf/9620

Licenses

Component	License	File
Model Weights	Apache 2.0	model/README.md Hub card
Dataset	CC BY-SA 4.0	HFREADME.md / Hub
Training Code	MIT	LICENSE / LICENSES.md

Component licenses: FiReCS CC BY 4.0, SentiTaglish CC BY 4.0, Amazon Apache 2.0, NewsPH public news text (no explicit license), TikTok CC BY-SA 4.0, adversarial CC BY-SA 4.0.

Related Resources

• Dataset: jjjardev/tagasenti — DOI 10.57967/hf/9620
• Code: github.com/jjjardev/tagasenti
• HiliSenti (sister project): jjjardev/hilisenti-v1 / jjjardev/hilisenti-v1-model (93.5% F1, Hiligaynon — structure reference for this card)
• Base model: xlm-roberta-large
• Paper notes (private, 9-day log v1–v6): paper_notes/COMPLETE_DOCUMENTATION.md (1305 lines), run_report.md (1575 lines) — distilled into this card

Contact

Jessie James T. Jarder — jj.jarder.dev@gmail.com
Issues: github.com/jjjardev/tagasenti/issues · HF discussion: huggingface.co/jjjardev/tagasenti_model/discussions