metadata
library_name: transformers
pipeline_tag: text-generation
tags:
- synthetic-data
- dpo
- gpqa
- reasoning
- alignment
- quantum
- neuroscience
- gloss-free
- data-efficient
base_model: Qwen/Qwen2.5-7B-Instruct
license: other
language:
- en
metrics:
- accuracy
datasets:
- TrueRunAI/TrueRun-Groove-v2.1-DPO
TrueRun-Groove-v2.1-7B
Qwen2.5-7B-Instruct fine-tuned on ~1,200 high-rigor synthetic DPO pairs (Groove v2.1).
Balanced quantum mechanics, neuroscience/BCI, alignment/game theory. Structural escalation for indefinite depth—no gloss decay.
Key Results (GPQA Diamond, 3 Seeds Mean)
| Benchmark | Questions | Baseline % | Groove Mean % | Delta | Notes |
|---|---|---|---|---|---|
| Full Diamond | 198 | 33.33% | 36.53% | +3.20% | Low variance (±0.58%) |
| Quantum Subset | 39 | 35.90% | 51.92% | +16.02% | Leading public targeted lift for 7B |
| Biology Subset | 19 | 36.84% | 52.63% | +15.79% | Strong transfer |
| Physics Subset | 86 | 51.16% | 42.25% | -8.91% | Targeted regression—next iter fix |
Leading data efficiency & domain-specific gains among public 7B fine-tunes.
License
Other (non-exclusive commercial/research use—dataset for sale on OpenDataBay; model weights public for testing/reproduction).
Usage
from transformers import pipeline
pipe = pipeline("text-generation", model="TrueRunAI/TrueRun-Groove-v2.1-7B")
pipe("Explain quantum entanglement simply but without losing rigor:", max_new_tokens=256)