🚦 tachyon-router-v1-mini

Turkish prompt complexity classifier for intelligent LLM routing.
Fine-tuned from dbmdz/electra-small-turkish-cased-discriminator — 13.7M parameters, 26.2MB (FP16).

The idea is simple: instead of letting users pick a model manually, this classifier reads the prompt and decides which model tier is appropriate — saving cost on simple queries while still routing complex ones to capable models.

---

This model was built with heavy AI assistance. The dataset was generated by Gemini 2.5 Flash, the training pipeline was written with Claude, and most architectural decisions were made through back-and-forth with AI tools — not from a formal ML background. That said, every output was reviewed, the metrics are real, and nothing was blindly copy-pasted. Call it AI-assisted rather than vibe-coded, but transparency feels right either way. Use with appropriate skepticism.

---

🏷️ Labels

Label	Name	Score Range	Description
0	trivial	0.05 – 0.22	Greetings, navigation, password reset, simple UI questions. No data retrieval needed.
1	simple	0.25 – 0.52	Single-step data lookup or filter. One table, one condition, no aggregation.
2	analysis	0.55 – 0.78	Multi-step reasoning: comparisons, summaries, trend analysis, cross-table queries.
3	complex	0.80 – 0.98	Forecasting, risk analysis, scenario modeling, root cause analysis, strategic recommendations.

---

📊 Metrics

Evaluated on a held-out validation set (n=210, balanced across classes).

Metric	Score
Accuracy	0.90
Macro F1	0.90
Macro Precision	0.90
Macro Recall	0.91

Per-class breakdown:

Class	Precision	Recall	F1
trivial	0.88	1.00	0.94
simple	0.97	0.88	0.92
analysis	0.89	0.84	0.86
complex	0.86	0.92	0.89

Confusion matrix:

              trivial  simple  analysis  complex
trivial          37       0        0        0
simple            5      57        2        1
analysis          0       2       47        7
complex           0       0        4       48

---

🚀 Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Squezilia/tachyon-router-v1-mini"
)

classifier("Bu ayki toplam ciro ne kadar?")
# [{'label': 'simple', 'score': 0.86}]

classifier("Önümüzdeki çeyrek için nakit akışı tahmini yap ve risk senaryolarını göster.")
# [{'label': 'complex', 'score': 0.76}]

With confidence threshold (recommended)

Low-confidence predictions are nudged to the next tier — safer for routing:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

LABEL_NAMES = {0: "trivial", 1: "simple", 2: "analysis", 3: "complex"}
CONF_THRESHOLD = 0.65

tokenizer = AutoTokenizer.from_pretrained("Squezilia/tachyon-router-v1-mini")
model = AutoModelForSequenceClassification.from_pretrained("Squezilia/tachyon-router-v1-mini")
model.eval()

def route(text: str) -> dict:
    enc = tokenizer(text, return_tensors="pt", max_length=128,
                    padding="max_length", truncation=True)
    with torch.no_grad():
        logits = model(**enc).logits
    probs = torch.softmax(logits, dim=1).squeeze()
    pred  = int(probs.argmax())
    conf  = float(probs[pred])

    # Bump up one tier if confidence is low
    if conf < CONF_THRESHOLD and pred < 3:
        pred += 1

    return {"label": pred, "label_name": LABEL_NAMES[pred], "confidence": conf}

route("Geçen ayla kıyasla gelir-gider özeti çıkar.")
# {'label': 2, 'label_name': 'analysis', 'confidence': 0.66}

---

🏗️ Training

Data

The training dataset was generated using Gemini 2.5 Flash with structured outputs, targeting Turkish business/ERP software prompts. It went through several iterations:

• v1: 1,104 examples — continuous regression labels (scores 0.0–1.0). The model learned to predict ~0.4 for everything (classic regression collapse on noisy labels).
• v2: Converted to 4-class classification, deduplicated, balanced to 250 examples per class with augmentation. Trained with dbmdz/distilbert-base-turkish-cased (80M params) → 0.77 accuracy.
• v3: Switched to dbmdz/electra-small-turkish-cased-discriminator (13.7M params), expanded dataset to ~1,539 examples including long/complex prompts → 0.88 accuracy.
• v4 (this model): Further expanded to ~1,539 examples with better domain coverage, tuned epochs and batch size → 0.90 accuracy. Exported to FP16 → 26.2MB.

Augmentation

To balance underrepresented classes, two light augmentation strategies were applied to the training set:

• Synonym replacement — domain-aware Turkish synonym swaps (e.g. göster → listele, karşılaştır → kıyasla)
• Word dropout — random word removal for sentences with 5+ words (p=0.10)

Hyperparameters

Parameter	Value
Base model	dbmdz/electra-small-turkish-cased-discriminator
Epochs	24 (early stopping at 15)
Batch size	16
Learning rate	2e-5
Warmup ratio	0.1
Weight decay	0.01
Max length	128
Loss	CrossEntropyLoss with class weights
Early stopping patience	3 (monitored: val macro F1)
Precision	FP16

---

🌍 Domain Coverage

The model was trained on prompts spanning these business domains:

• Muhasebe & Finans — gelir, gider, fatura, vergi, bilanço
• Stok & Depo — ürün, stok, sipariş, tedarik
• Satış & Müşteri — ciro, müşteri, kampanya, tahsilat
• İnsan Kaynakları — personel, maaş, izin, performans
• Üretim & Operasyon — vardiya, maliyet, kapasite, fire
• Sistem & Kullanıcı — şifre, yetki, ayar, bildirim

---

⚠️ Limitations

• Trained exclusively on Turkish business/ERP prompts — not suitable for general-purpose or non-Turkish text.
• The analysis / complex boundary is the hardest to classify (semantically close). The confidence threshold nudge helps mitigate misrouting.
• Dataset was synthetically generated — real-world distribution may differ.

---

📦 Model Files

File	Size	Description
model.safetensors	55 MB	FP32 weights
model_fp16.safetensors	27.5 MB	FP16 weights (recommended)
model.onnx	55.2 MB	ONNX export (FP32)
model_fp16.onnx	27.8 MB	ONNX export (FP16, for browser/edge)
config.json	1.25 kB	Model config
config_fp16.json	1.25 kB	FP16 model config
config_onnx.json	980 B	ONNX model config
tokenizer.json	755 kB	Tokenizer
tokenizer_config.json	342 B	Tokenizer config

---

Built for a personal ERP assistant project. Feedback welcome.