← Back to Home
HELIX Research and Benchmark Materials
This page consolidates investor and technical diligence artifacts, including measured benchmark outputs, reproducibility notes, checkpoint coverage, and direct paper access.
Current Paper
Read the latest arXiv preprint and contact us for full patent portfolio review packets.
| Benchmark | Temperature | Metric | Value | Steering Rate | Source Artifact |
|---|---|---|---|---|---|
| GSM8K | 1.0 | Accuracy | 91.80% | 1.72% | ICML_Experiments/results/gsm8k_Cortex_(4-bit_Granite)_t1.0_final.json |
| GSM8K | 3.0 | Accuracy | 88.84% | 2.44% | ICML_Experiments/results/gsm8k_Cortex_(4-bit_Granite)_t3.0_final.json |
| MMLU | 1.0 | Accuracy | 74.30% | 0.02% | ICML_Experiments/results/mmlu_Cortex_(4-bit_Granite)_t1.0_final.json |
| MMLU | 3.0 | Accuracy | 72.49% | 0.40% | ICML_Experiments/results/mmlu_Cortex_(4-bit_Granite)_t3.0_final.json |
| HumanEval | 1.0 | Pass@1 | 81.10% | 0.42% | ICML_Experiments/results/humaneval_Cortex_(4-bit_Granite)_t1.0_final.json |
| HumanEval | 2.5 | Pass@1 | 80.49% | 0.33% | ICML_Experiments/results/humaneval_Cortex_(4-bit_Granite)_t2.5_final.json |
Artifact Coverage
- • Final JSON result files: 43
- • Checkpoint JSON files: 473
- • Benchmark sets represented: GSM8K, MMLU, HumanEval, TruthfulQA
Data Integrity Notes
- • Values shown above are from measured `summary` fields in final artifacts.
- • Telemetry coverage in sampled finals is 1.0 (100%).
- • Full raw files are available via diligence package on request.
Downloadable Materials
ICML Experiment Suite README
Full experiment orchestration, dependencies, and reproducibility workflow.
A4 Reproducibility Summary
Hyperparameters, manifold protocol, and telemetry checklist.
OptSetting v1 Analysis
Steering-collapse mitigation analysis and temperature trade-offs.
ICML Results Summary (JSON)
Curated benchmark metrics from final result artifacts.
Checkpoint Catalog Summary (JSON)
Final artifact and checkpoint counts across benchmarks.