fluvial-gemma-r13-merged

TranslateGemma R13 — Fluvial Constellations Term Translator

A fine-tuned Gemma 2 2B-it model that translates Fluvial Constellations (Constelaciones Fluviales) therapy terminology from Spanish to English.

100% accuracy on all 106 glossary terms (Glossary v3.0) and 58/58 on real class transcript tests.

What is Fluvial Constellations?

Fluvial Constellations (Constelaciones Fluviales) is a therapeutic method developed by Juan Carlos Arias Quintana, inspired by Bert Hellinger's Family Constellations. It is a systemic model performed in water using figurines in an inflatable pool, with its own specialized terminology that does not have standard English translations.

This model was trained to provide consistent, accurate English translations of the 106 official terms in the discipline's bilingual glossary.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "eregon/fluvial-gemma-r13-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

def translate(term):
    prompt = (
        "<start_of_turn>user\n"
        f"Translate this Fluvial Constellations therapy term to English: {term}"
        "<end_of_turn>\n"
        "<start_of_turn>model\n"
    )
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        out = model.generate(**inputs, max_new_tokens=64, do_sample=False)
    response = tokenizer.decode(out[0], skip_special_tokens=False)
    response = response.split("<start_of_turn>model\n")[-1]
    return response.replace("<end_of_turn>", "").replace("<eos>", "").strip()

# Examples
print(translate("Constelaciones Fluviales"))  # → Fluvial Constellations
print(translate("Órdenes del Amor"))          # → Orders of Love
print(translate("Implicación"))               # → Entanglement
print(translate("Buen Lugar"))                # → Proper Place
print(translate("Tirada de Diagnóstico"))     # → Diagnosis Throw

With 4-bit quantization (lower VRAM)

from transformers import BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)
model = AutoModelForCausalLM.from_pretrained(
    model_id, quantization_config=bnb_config, device_map="auto"
)

Complete Glossary (106 terms)

Training Details

Model Lineage

google/gemma-2-2b-it
  → R7-merged → R8 LoRA (91.67%)
    → R8-merged → R11 LoRA (95.37%)
      → R11-merged → R12 LoRA (100%, Glossary v2.1)
        → R12-merged → R13 LoRA (100%, Glossary v3.0) → R13-merged  ← this model

R13 Training Configuration

R13 Dataset

325 training examples across 4 difficulty tiers:

• Tier 1 (15 reps, 3 terms): Hardest changes — acronym remappings (COM→COL, LEO→PEO) and new term (Mirada Rebasada)
• Tier 2 (10 reps, 15 terms): Modified translations (e.g., "Look of Love", "Remote Constellation", "Corporate Hydrosystemics")
• Tier 3 (8 reps, 6 terms): Simplified terms (removed / alternatives, e.g., "Pool" instead of "Pool / Swimming Pool")
• Tier 4 (1 rep, 82 terms): Reinforcement for unchanged terms to prevent regression

Training Results

Evaluation Results

Glossary Version History

• v3.0 (R13, current): 106 terms. Simplified slash alternatives, English acronyms (COL, PEO), Hellinger-standard terminology (Entanglement, Proper Place)
• v2.1 (R12): 108 terms. Added transcript-derived terms
• v1.0 (R7-R11): Initial glossary builds

Intended Use

This model is designed for:

• Real-time translation of Fluvial Constellations terminology during classes and sessions
• Subtitle generation for Spanish-language instructional videos
• Supporting bilingual documentation of the methodology

Limitations

• Trained specifically for Fluvial Constellations terminology — not a general-purpose translator
• Best results with the exact prompt format shown above
• Some terms (e.g., "Campo", "Orden") are common Spanish words; the model translates them in the therapeutic context

Hardware

• Trained on NVIDIA RTX 2080 (8 GB VRAM)
• Inference: ~1.6 GB (4-bit) or ~4.9 GB (bf16)