GetGenetica/jearx-gemma4-e4b-text2sql-sota

JEA-RX Gemma 4 E4B Text2SQL — SOTA

Best available checkpoint for JEA-RX text-to-SQL generation. Trained April 2026.

Base: google/gemma-4-E4B-it + LoRA adapters Training data: 38,887 exec-verified SQL pairs from GetGenetica/jearx-text2sql-training-data Post-training: 200 GRPO RL steps with composite neurosymbolic reward

This model is the base checkpoint for the JEA-RX v3.0 ablation study (GetGenetica/jearx-ablation-v3-*).

Honest benchmarks

Benchmark	Metric	Score	Notes
BIRD holdout (50 rows)	exec_acc (real SQLite)	0.610	Gemma 3 4B baseline — this model expected to exceed
BIRD holdout (50 rows)	exec_acc (real SQLite)	TBD	Run ablation_full to measure
Spider 1034	exec_acc	TBD	Pending evaluation run

Note: exec_acc reported here uses real SQLite execution. Prefix-validity scores (which give 1.0) are NOT reported.

Architecture

Property	Value
Base	Gemma 4 E4B Instruct (3.77B params)
LoRA rank	r=8, α=32
LoRA targets	q_proj, k_proj, v_proj, o_proj
Training examples	38,887 (exec_acc=1.0 only, curated)
Data mix	60% curated_clean + 30% BIRD + 10% Tray golden
GRPO RL steps	200 post-SFT
GRPO reward	0.4×graph + 0.4×cispo + 0.2×format (v2.9 weights)
Platform	JEA-RX neurosymbolic routing (Tsetlin + Clifford Algebra)

v3.0 Upgrade path

This model feeds into the SOTA v3.0 training which adds:

Component	Role	Weight
NCD schema grounding	Identifier-level overlap (schema idents ∩ SQL idents)	0.25
ChromatinHDC curriculum	1024-dim bipolar HDC novelty upsampling (2× for novel > 0.3)	data
B-spline mini-KAN	5→3→1 interpretable reward network, frozen after 50 steps	0.08
Hallucination kill-switch	ncd_schema < 0.05 AND exec_sim < 0.05 → reward = 0.0	—

See GetGenetica/jearx-ablation-v3-full for the component ablation study.

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "google/gemma-4-E4B-it", torch_dtype=torch.bfloat16, device_map="auto"
)
tok = AutoTokenizer.from_pretrained("google/gemma-4-E4B-it")
model = PeftModel.from_pretrained(base, "GetGenetica/jearx-gemma4-e4b-text2sql-sota")
model.eval()

prompt = "[dialect:sqlite] [question] Show total revenue by location for the last 7 days [/question]\n\n### SQL:\n"
inputs = tok(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=150, do_sample=False)
sql = tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(sql)

Prompt format

[dialect:sqlite] [schema] CREATE TABLE orders (id INT, amount FLOAT, location TEXT, date DATE);

[question] Show total revenue by location for the last 7 days [/question]

### SQL:

Supported dialect tags: [dialect:sqlite], [dialect:snowflake], [dialect:postgresql], [dialect:cypher]

Training data

From GetGenetica/jearx-text2sql-training-data:

File	Rows	Source
curated_sql_clean_v1.jsonl	38,887	17 sources, 97.2% exec_verified, 49% paraphrases
bird_training_bundle_v2.jsonl	8,505	BIRD-SQL training set
tray_golden_v3.jsonl	97	Tray POS domain golden pairs

FORGE integration

This model is served via FORGE (/sql endpoint) with Kolmogorov complexity scoring and Tsetlin Machine routing:

curl http://localhost:3030/sql \
  -H "Content-Type: application/json" \
  -d '{"question": "Show daily revenue", "dialect": "sqlite"}'

Response includes sql, verdict (PASS/FAIL/UNAVAILABLE), reward (0–1), and routed_by.

Platform

JEA-RX (Joint Embedding Addressable — Relational eXpert) platform components:

• FORGE: Rust/WASM agent orchestration runtime with Tsetlin + Clifford routing
• MCP server: serving/forge_mcp_server.py — exposes 10+ tools to any MCP client
• Mechanistic interpretability: 6-layer pipeline trace (L0–L5) in FORGE UI
• NIST AI RMF 1.0: compliance endpoints at /api/stateless/nist-status

Ablation study

Run hf_jearx_sota_v3_ablation.py with ABLATION_MODE to compare components:

Mode	Description	Expected result
full	All v3 components (control)	Highest mean reward
no_ncd	No schema grounding	Hallucination rate increases
no_hdc	No novelty curriculum	Fewer rare SQL patterns learned
no_kan	No B-spline KAN reward	Slight reward reduction
v2_compat	v2.9 reward structure (baseline)	Lower mean reward