ravinarayanan/Gemma4-E2b-SOQL-QLORA

Gemma 4 E2B — SOQL Query Generator (QLoRA, Q4_K_M GGUF)

Fine-tuned from google/gemma-4-2b-it to translate natural-language questions into SOQL (Salesforce Object Query Language) queries.

Note: The model is trained and prompted to return bare SOQL, but like all small LLMs it may occasionally include reasoning traces (<think>...</think> blocks), brief explanations, or off-script text. Always validate the output and strip non-SOQL content before execution. The reference orchestrator (orchestrator/app.py) handles this automatically.

---

Model details

Property	Value
Base model	unsloth/gemma-4-E2B-it (4-bit, Unsloth day-0 quant)
Fine-tune method	QLoRA — LoRA r=16, alpha=16
LoRA target modules	q/k/v/o_proj + gate/up/down_proj
Training hardware	A100 GPU
Training steps	250 (early stop; val loss 0.119)
Dataset	11,040 synthetic NL→SOQL pairs, 17 SOQL pattern types
Quantization	Q4_K_M GGUF (~1.5 GB)
Context window	2048 tokens

---

What it does

Given a Salesforce object schema + a natural-language question, the model returns a SOQL SELECT query. It was trained on fictional object/field names (Vehicle__c, Driver__c, etc.) to learn SOQL syntax — not any specific org's schema. Inject your real org's schema in the user message at inference time.

SOQL patterns covered

Simple SELECT · WHERE filters · ORDER BY · LIMIT · COUNT/aggregate · GROUP BY · HAVING · date literals · multi-field · relationship queries · subquery IN filters · parent field traversal · child subqueries · LIKE · NULL checks · boolean logic · semi-joins

---

Usage with Ollama (recommended)

# Pull directly from this repo
ollama pull hf.co/ravinarayanan/gemma-2b-soql-qlora

# Or create from the bundled Modelfile
ollama create soql-gemma4 -f Modelfile
ollama run soql-gemma4

Example prompt

Object: Vehicle__c
Fields: Name, Make__c, Model__c, Year__c, Price__c, FuelType__c, IsAvailable__c
Relationships: Dealership__c (lookup), ServiceRecords__r (child)

Write a SOQL query to find electric vehicles priced under 25000 that are available

Example output

The model typically returns bare SOQL, but may prefix it with a <think> block. Strip everything before and including </think> if present:

SELECT Name, Make__c, Model__c, Price__c
FROM Vehicle__c
WHERE FuelType__c = 'Electric'
AND Price__c < 25000
AND IsAvailable__c = true

Parsing the output in Python

import re

def extract_soql(raw: str) -> str:
    # Strip <think>...</think> block if present
    raw = re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL).strip()
    # Take only the first SELECT statement
    match = re.search(r"(SELECT\b.*)", raw, re.IGNORECASE | re.DOTALL)
    return match.group(1).strip() if match else raw.strip()

---

System prompt used at inference

You are a SOQL query generator. Given a Salesforce schema and a question,
output ONLY the bare SOQL query. No explanation, no markdown, no extra text,
no reasoning.

---

Limitations

• Output not guaranteed to be pure SOQL — the model may include reasoning traces or brief text; validate and strip before execution
• Trained on synthetic data — always test generated queries against your org
• Only SELECT queries — DML (INSERT/UPDATE/DELETE) is intentionally excluded from training
• Schema must be provided in the user message; the model has no live Salesforce access
• 2B parameter model — complex nested subqueries may need a retry or prompt adjustment

---

Training data

11,040 synthetic training examples generated from a fictional Salesforce schema (4 objects: Vehicle__c, Driver__c, Dealership__c, ServiceRecord__c) covering 17 SOQL pattern types. Dataset was generated with a capability-aware template generator to ensure pattern diversity and avoid duplicates.

---

License

This model is derived from Gemma 4 and is subject to the Gemma Terms of Use.