davidburhans/gevva-decisions

⚡ Gevva Decisions: Complete Training Curriculum & Benchmark Suite

Official training mixtures, committee-verified synthetic datasets, and held-out benchmarks used to train and evaluate the Gevva family of large-context (128K), multimodal System 1 decision engines (#1 Global on JevBench: 77.54 Composite Score).

• Associated Models:
  • Flagship: davidburhans/gevva-e2b (Default main branch)
  • Multimodal: davidburhans/gevva-e2b (branch multimodal)
• Codebase & Reproduction Scripts: https://github.com/davidburhans/gevva

---

📦 Dataset Configurations & Splits

Configuration	File	Records	Description	Primary Role
flagship	data/flagship_train.jsonl	145,000+	Master Phase-3 enriched curriculum	Reproduces #1 Global JevBench Champion (77.54)
multimodal	data/multimodal_train.jsonl	65,472	Phase-4 multimodal + anchor replay	Trains vision grounding (88.8% accuracy, 96.4% tables)
sdk_decisions	data/sdk_decisions_train.jsonl / val	8,515	4-judge committee-verified SDK pairs	Tool routing, candidate reranking, rubric grading
remediation	data/weak_family_remediation.jsonl	7,600	Targeted weak-family reasoning pairs	Complex policy precedence, arithmetic, multi-hop
benchmarks	data/test_decisions.jsonl	2,947	Held-out decision benchmark	Generalization testing & downstream decision scoring
benchmarks	data/test_multimodal.jsonl	249	Held-out visual grounding benchmark	Evaluates tables, charts, invoices, and spatial scenes

---

🛠️ How to Load via Hugging Face datasets

from datasets import load_dataset

# 1. Load the flagship master training mixture
flagship_ds = load_dataset("davidburhans/gevva-decisions", "flagship")
print(flagship_ds["train"][0])

# 2. Load SDK-parity decision pairs (tool routing, search reranking, grading)
sdk_ds = load_dataset("davidburhans/gevva-decisions", "sdk_decisions")
print(f"SDK Train: {len(sdk_ds['train'])}, SDK Val: {len(sdk_ds['validation'])}")

# 3. Load held-out test benchmarks
benchmarks = load_dataset("davidburhans/gevva-decisions", "benchmarks")
print(f"Decisions Test: {len(benchmarks['test_decisions'])}, Multimodal Test: {len(benchmarks['test_multimodal'])}")

---

🔬 Dataset Schema & Record Format

All records follow the unified cross-encoder / decision format:

{
  "premise": "Customer account balance is $450.00 with pending refund request for $45.00.",
  "hypothesis": "Process a refund of $45.00 to original payment method.",
  "label": 1,
  "source": "sdk_synthetic_tool_routing",
  "group_id": 1042,
  "language": "en"
}

Label Conventions:

Following the standard ModernCE / openjev specification:

• 0: Contradiction (Reject candidate / refuted claim)
• 1: Entailment (Select candidate / verified claim)
• 2: Neutral (Unverifiable / irrelevant candidate)

---

👥 Authors & Co-Authorship

• Dave Burhans — Lead Author & Architecture
• Gemini 3.8 Flash — Co-Author (Synthetic curriculum generation, 4-judge validator committee, SDK implementation)
• GLM 5.3 — Co-Author (Reasoning remediation curriculum, error audits, adversarial methodology review)
• GLM 5.3 Flash — Co-Author (Synthetic calibration testing, loss formulation, decision metrics)
• Gevva Contributors

---

🙏 Model & Quant Provider Acknowledgments

The synthetic decision generation and multi-judge validation pipeline relies on high-throughput local inference via llama-server (port 8080). We gratefully acknowledge and credit the following foundation models, creators, and quantization providers:

Foundation Models:

• Qwen 3.6 27B (Qwen/Qwen3.6-27B-MTP) & Qwen 3.8 Flash Next (Qwen/Qwen3.8-Flash-Next MoE) by the Qwen Team / Alibaba Cloud.
• DeepSeek V4 Flash (deepseek-ai/DeepSeek-V4-Flash) by DeepSeek AI.
• Gemma 4 31B (google/gemma-4-31B-it) by Google DeepMind.

Quantization & GGUF Creators:

• Unsloth AI (@unsloth):
  • unsloth/Qwen3.6-27B-MTP-GGUF (Qwen3.6-27B-UD-Q4_K_XL.gguf with draft MTP)
  • unsloth/Qwen3.8-Flash-Next-GGUF (UD-IQ4_XS and shared draft MTP)
  • unsloth/DeepSeek-V4-Flash-0731-GGUF (UD-IQ4_NL)
• ISTA-DASLab (@ISTA-DASLab):
  • ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF (IQ3_XXS using Generalized Second-Order Quantization & Residual Compensation Optimization)
• Bullerwins (@bullerwins):
  • bullerwins/DeepSeek-V4-Flash-0731-GGUF (DSPARK, Q8_0)
• Google:
  • google/gemma-4-31B-it-qat-q4_0-gguf

---

👥 Attribution, Credits & Foundation Models

The synthetic data generation, weak-family remediation, and multi-judge validation pipelines for Gevva Decisions leverage state-of-the-art open models and quantized runtimes:

• Base Models:
  • Qwen 3.6 27B MTP and Qwen 3.8 Flash Next, developed by the Qwen Team at Alibaba Cloud (Qwen).
  • Google Gemma 4 family (google/gemma-4-E2B-it), developed by Google DeepMind.
• Quantization & GGUF Providers:
  • The high-throughput GGUF quants served via llama-server on port 8080 were created and published by Unsloth AI (unsloth):
    • unsloth/Qwen3.6-27B-MTP-GGUF (Quant: Q4_K_XL)
    • unsloth/Qwen3.8-Flash-Next-GGUF (Quant: IQ4_XS)
  • Inference execution powered by the open-source llama.cpp runtime.

---

📜 Citation & License

This dataset is released under the Apache 2.0 License.

@dataset{gevva_decisions_2026,
  author = {Burhans, Dave and {Gemini 3.8 Flash} and {GLM 5.3} and {GLM 5.3 Flash} and Contributors},
  title = {Gevva Decisions: Master Training Curriculum & Benchmark Suite for Multimodal System 1 Engines},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/davidburhans/gevva-decisions}
}