---
library_name: transformers.js
base_model: ravinarayanan/gemma-2b-soql-qlora
tags:
  - soql
  - salesforce
  - onnx
  - gemma
---

# Gemma 4 E2B SOQL (ONNX, text-only)

Merged from [`ravinarayanan/gemma-2b-soql-qlora`](https://huggingface.co/ravinarayanan/gemma-2b-soql-qlora) (QLoRA adapter) onto `unsloth/gemma-4-E2B-it`, stripped to text-only, exported to ONNX via the transformers.js community converter.

## Validation results

18 held-out NL->SOQL prompts from `dataset_val_raw.jsonl` (never seen during training), exact-match after whitespace/case normalization. Generated with `{ "role": "system", "content": "You are a helpful Salesforce AI assistant. Answer general questions using your own knowledge. When asked to write a SOQL query, reason through the schema and requirements, then produce the correct SOQL query." }` and greedy decoding.

| Variant | vs. ground truth | vs. bf16 baseline |
|---|---|---|
| baseline (merged PyTorch, bf16) | 0.0% | -- |

**Note:** ONNX-side accuracy could not be measured in the environment this was published from (`optimum.onnxruntime` unavailable/incompatible). Published quant `q4` is an untested default, not a validated pick -- confirm behavior with the Node.js smoke test (step 9 of the conversion notebook) before relying on it.


Published quant: **q4** (smallest quant that did not regress vs. baseline).


Validated on 2026-07-04.