Pythia 1.4B -- jBlaze Reasoning Enhanced
A reasoning-enhanced version of EleutherAI's Pythia 1.4B, modified using jBlaze behavioral weight editing. No gradient-based training was performed. The model has the same parameter count as vanilla Pythia 1.4B. The only difference is targeted edits to weight matrices using jBlaze.
What changed
jBlaze applies behavioral modifications directly to model weight matrices. This produces targeted changes in the model's decision behavior without retraining. The reasoning-enhanced version applies two complementary behavioral modifications across all 24 transformer layers.
No fine-tuning. No LoRA. No gradient descent. No additional training data.
Benchmark results
Evaluated on a 100-question held-out reasoning benchmark across 5 categories. These questions were never used during the configuration search. Evaluation uses logit-level multiple-choice scoring (the same methodology as MMLU).
| Model | Overall | Logic | Math | Sequences | Verbal | Causal | Perplexity |
|---|---|---|---|---|---|---|---|
| Raw Pythia 1.4B | 11% | 15% | 10% | 10% | 10% | 10% | -- |
| DNP base (before reasoning edits) | 7% | 20% | 10% | 5% | 0% | 0% | 2.64 |
| jBlaze Reasoning | 56% | 55% | 60% | 30% | 60% | 75% | 3.20 |
Overall held-out reasoning accuracy increased from 7% to 56% -- an 8x result, equivalent to a 700% relative increase in benchmark accuracy.
The gain is broad -- every category improved. Causal reasoning moved from 0% to 75%. Verbal reasoning from 0% to 60%. Math from 10% to 60%.
Perplexity increased from 2.64 to 3.20, indicating a modest distributional shift rather than catastrophic degradation.
Important context
- This is a 1.4 billion parameter model. It does not produce high-quality free-form text regardless of reasoning edits. The improvement is measured at the logit/decision level, which is how standard benchmarks (MMLU, ARC, etc.) evaluate reasoning.
- The benchmark contains 100 questions across 5 categories. Larger independent benchmarks would strengthen or weaken these results -- that is why we are publishing the model and evaluation script.
- "8x higher accuracy on this benchmark" is accurate. "8x smarter" is not what we are claiming.
- The base model was first processed with DNP (Direct Neural Programming) for knowledge injection before reasoning blazes were applied. The 7% baseline reflects the DNP model. Raw Pythia 1.4B scores 11% on the same benchmark -- both are near-random for reasoning.
Test it yourself
The evaluation script is included in this repository. Compare any model against this benchmark:
# Test the jBlaze reasoning-enhanced model
python eval_reasoning.py --model ApolloRaines/Pythia-1.4B-jBlaze-Reasoning
# Test vanilla Pythia 1.4B
python eval_reasoning.py --model EleutherAI/pythia-1.4b
# Test any other model
python eval_reasoning.py --model <your-model-path-or-hf-id>
Requirements: torch, transformers
How to use
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("ApolloRaines/Pythia-1.4B-jBlaze-Reasoning")
tokenizer = AutoTokenizer.from_pretrained("ApolloRaines/Pythia-1.4B-jBlaze-Reasoning")
What is jBlaze?
jBlaze is the behavioral weight editing compiler built by SAIQL. It applies targeted modifications to model weight matrices, changing specific behavioral traits without retraining.
The technology behind jBlaze is proprietary. The method is not published. The results are.
Released models: huggingface.co/ApolloRaines
Technical details
- Base model: EleutherAI/pythia-1.4b (1.4B parameters, 24 layers, GPT-NeoX architecture)
- Modification method: jBlaze weight editing (no gradients)
- Parameter count: Unchanged (1,414,647,808)
- Gradient-training data used: None
- Compute cost: Under 2 minutes on an NVIDIA RTX 3090
Citation
@misc{raines2026jblaze-reasoning,
title={jBlaze Reasoning Enhancement: Improving Held-Out Reasoning via Behavioral Weight Editing},
author={Apollo Raines},
year={2026},
url={https://huggingface.co/ApolloRaines/Pythia-1.4B-jBlaze-Reasoning}
}
A Note on Our Released Models
Most of our publicly released models are intentionally left at partial strength. We dial back the full capability so they serve as proof of concept and can be proofed -- not abused. The point is to show what's possible, not to hand it out at full power. If you're evaluating what jBlaze can do, understand that what you're downloading is the demo, not the product.