GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ravinarayanan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ravinarayanan
Model Tree
google/gemma-2b-it
Input Modalities
Output Modalities
Supported Functionality
| Property | Value |
|---|---|
| Base Model | Google Gemma 2B (unsloth/gemma-4-E2B-it) |
| Parameters | ~2 Billion |
| Fine-tune Method | QLoRA — 4-bit quantized, LoRA r=16, α=16 |
| LoRA Target Modules | q/k/v/o_proj, gate/up/down_proj |
| Training Examples | 2,024 train / 176 validation |
| SOQL Pattern Types | 17 (simple filters → nested subqueries) |
| Max Sequence Length | 1,024 tokens |
| Training Hardware | Google Colab T4 GPU |
| Training Time | ~20 minutes |
| Final Train Loss | 0.0327 |
| Final Val Loss | 0.0676 |
| Framework | Unsloth + TRL SFTTrainer |
| License | Apache 2.0 |
Given a Salesforce object schema and a natural-language question, the model outputs only the raw SOQL query — no markdown fences, no explanation, no reasoning trace.
text
Schema: Object: Vehicle__c. Fields: Name, Make__c, FuelType__c, Price__c, Color__c.Question: Show electric vehicles priced under 25000→ SELECT Id, Name, Make__c, Price__c FROM Vehicle__c WHERE FuelType__c = 'Electric' AND Price__c < 25000Why provide schema at inference time?
The model was trained on fictional Salesforce objects (Vehicle__c, Driver__c, etc.) to learn SOQL syntax, not any specific org's schema. At inference time, inject the real object's schema in the user message — matching the training format exactly.
| # | Pattern | Example Clause |
|---|---|---|
| 1 | Basic equality | WHERE FuelType__c = 'Electric' |
| 2 | Not-equal | WHERE Color__c != 'Red' |
| 3 | Numeric compare | WHERE Price__c < 20000 |
| 4 | ORDER BY + LIMIT | ORDER BY Price__c DESC LIMIT 5 |
bash
pip install transformers torch acceleratepython
from transformers import AutoModelForCausalLM, AutoTokenizerimport torch model_id = "ravinarayanan/soql-gemma4-lora"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.float16, device_map="auto",)model.eval() SYSTEM_PROMPT = ( "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.") def generate_soql(schema: str, question: str) -> str: messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": f"{schema}\n\nQuestion: {question}"}, ] # enable_thinking=False is REQUIRED — prevents Gemma 4 reasoning trace prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False ) inputs = tokenizer(prompt, return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=120, do_sample=False) decoded = tokenizer.decode(outputs[0], skip_special_tokens=True) return decoded[len(tokenizer.decode(inputs["input_ids"][0], skip_special_tokens=True)):].strip() # Exampleschema = "Object: Vehicle__c. Fields: Name, Make__c, FuelType__c, Price__c, Color__c."soql = generate_soql(schema, "Find electric vehicles priced under 25000")print(soql)# SELECT Id, Name, Make__c, Price__c FROM Vehicle__c WHERE FuelType__c = 'Electric' AND Price__c < 25000python
from unsloth import FastLanguageModel model, tokenizer = FastLanguageModel.from_pretrained( "ravinarayanan/soql-gemma4-lora", max_seq_length=1024, dtype=None, load_in_4bit=True,)FastLanguageModel.for_inference(model)python
from langchain_community.llms import HuggingFacePipelinefrom transformers import pipeline pipe = pipeline( "text-generation", model=model, tokenizer=tokenizer, max_new_tokens=120, do_sample=False,)llm = HuggingFacePipeline(pipeline=pipe)python
SFTConfig( per_device_train_batch_size = 8, gradient_accumulation_steps = 2, num_train_epochs = 3, learning_rate = 2e-4, lr_scheduler_type = "cosine", warmup_ratio = 0.05, optim = "adamw_8bit", max_seq_length = 1024,)| Step | Train Loss | Val Loss |
|---|---|---|
| 50 | 0.3821 | 0.4103 |
| 100 | 0.1820 | 0.1944 |
| 150 | 0.1122 | 0.1187 |
| 200 | 0.0821 | 0.0893 |
| 250 | 0.0612 | 0.0741 |
| 300 | 0.0447 | 0.0698 |
⚠️
enable_thinking=Falseis mandatory.
Gemma 4 has a built-in thinking/reasoning mode. Always passenable_thinking=Falsetoapply_chat_template— otherwise the model outputs a reasoning trace before the SOQL.
⚠️ Validate output before executing.
This model was trained exclusively on SELECT queries. Reject any output not beginning withSELECTbefore sending to Salesforce.
ℹ️ Schema injection required.
The model expects the object schema in the user message. Without it, query quality degrades significantly.
bibtex
@misc{soql-gemma4-lora-2025, author = {Ravi Narayanan}, title = {soql-gemma4-lora: Gemma 2B Fine-Tuned for Salesforce SOQL Generation}, year = {2025}, publisher = {Hugging Face}, url = {https://huggingface.co/ravinarayanan/soql-gemma4-lora}}| 5 |
| Multi-field SELECT |
SELECT Name, Make__c, Year__c, Price__c |
| 6 | COUNT aggregate | SELECT COUNT() FROM Driver__c |
| 7 | SUM / AVG / MIN / MAX | SELECT AVG(Price__c) FROM Vehicle__c |
| 8 | GROUP BY | GROUP BY Make__c |
| 9 | HAVING | HAVING COUNT() > 3 |
| 10 | IN / NOT IN | WHERE Color__c IN ('Red', 'Blue', 'Green') |
| 11 | LIKE | WHERE Name LIKE '%Toyota%' |
| 12 | Date literals | WHERE ServiceDate__c = LAST_MONTH |
| 13 | Boolean | WHERE IsActive__c = true |
| 14 | NULL check | WHERE Driver__c = null |
| 15 | Parent traversal (child→parent) | SELECT Dealership__r.Name FROM Vehicle__c |
| 16 | Subquery (parent→child) | SELECT (SELECT Name FROM Vehicles__r) FROM Driver__c |
| 17 | Cross-object WHERE | WHERE Dealership__r.Region__c = 'North' |
| 381 | 0.0327 | 0.0676 |