IgnisCogitationis/quantum-like-attention-framework-1.3b-untuned-validation

Quantum Like Attention Framework (Q.L.A.F) 1.3b untuned

This repository contains the model checkpoints, downstream evaluation scores, and pretraining convergence logs for the Quantum Like Attention Framework (Q.L.A.F) 1.3B configuration.

---

Key Specifications & Architecture

• Model Name: Q.L.A.F 1.3b untuned (Quantum Like Attention Framework - Hybrid Architecture)
• Parameters: 1.3B parameters total configuration (327M active parameter student subset)
• Layer Count: 12 layers (Hybrid: 9 QLAF recurrent layers + 3 FlashAttention layers, interleaved every 4th layer)
• Architecture Distribution: 75% QLAF Recurrent Layers / 25% FlashAttention Layers (every 4th layer: index 3, 7, 11)
• Recurrent State Size (S): 24
• Recurrent State Rank (R): 192
• Hidden Size (H): 2048
• Intermediate FFN Size (F): 5632
• Recurrent Normalization: Parameter-free Recurrent RMSNorm integrated into recurrent updates.
• Attention Mechanism: Native FlashAttention-2 causal scaled dot-product attention on interleaved layers.

---

Pretraining & Training Method

Pretraining is executed under the Odyssey Route-B Unified Engine with PyTorch Distributed Data Parallel (DDP) scaling support.

• Data Mixture: 35% FineWeb-Edu, 30% SlimPajama-6B, 15% GitHub Code, 10% STEM/ArXiv, 5% DCLM, and 5% Synthetic Multi-Hop Stories (with 10% retrieval batches).
• Curriculum Context Growth: Exponential context scaling starting from 4,096 tokens up to 65,536 tokens over the campaign steps.
• Global Batch Size: Targeted at 65,536 tokens (using dynamic gradient accumulation steps based on curriculum length).
• Optimizer: AdamW (beta1=0.9, beta2=0.999, weight decay = 0.01)
• Learning Rate Schedule: Cosine decay with 6,000 warmup steps, scaling from 0 up to a peak of 3e-4 and decaying to 3e-5.
• Routing Entropy Exploration: To prevent the AION router from getting stuck in local minima early in training, a routing entropy loss is added to the training objective during the first 20,000 steps with a linear decay weight: $$\text{entropy_weight} = 0.03 \times \left(1.0 - \frac{\text{step}}{20000}\right)$$ This forces the router to explore uniform row distributions, preventing it from locking onto random rows early and guaranteeing convergence to the correct state updates for secret key retrieval.

---

Downstream Evaluation Results (3 Seeds)

Evaluations are conducted across three independent seeds (42, 100, 2026). Accuracies are reported on MMLU (500 validation questions), HumanEval (164 coding tasks), and the 65,000-token Needle-in-a-Haystack retrieval benchmark.

Seed	Final Loss	Perplexity	MMLU Score	HumanEval (Assertions Passed)	Needle-in-a-Haystack (65k Context)	Status
Seed 42	Pending	Pending	Pending	Pending	Pending	Running
Seed 100	Pending	Pending	Pending	Pending	Pending	Pending
Seed 2026	Pending	Pending	Pending	Pending	Pending	Pending

---

Licenses

This model card, the accompanying code, configs, and model weights are licensed under the custom personal/non-commercial license:

1. Model Weights, Code, and Configs: Licensed under a custom non-commercial license. Personal use is fully allowed. Commercial use is not allowed without prior written permission (see LICENSE.md).
2. SlimPajama Datasets: Subject to the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
3. FineWeb-Edu: Licensed under the Open Data Commons Attribution License (ODC-By) v1.0.

---

Complete Reproducibility Guide

To reproduce the pretraining run and evaluations from scratch:

1. Clone and Install: Ensure PyTorch 2.4+ and Triton 3.0+ are installed in your environment.

   pip install torch triton transformers datasets huggingface_hub
   

2. Run Pretraining (DDP Multi-GPU configuration): Launch the training script on an 8-GPU node:

   torchrun --nproc_per_node=8 run_1_3b_scaling_proof.py \
     --seed 42 \
     --steps 75000 \
     --S 24 \
     --R 192 \
     --num_layers 12 \
     --retrieval_ratio 0.10 \
     --retrieval_loss_weight 3.0 \
     --lr 3e-4 \
     --warmup 6000
   

3. Run Downstream Evaluations: Execute the evaluation pipeline on the completed checkpoint:

   python scripts/run_validation_entrypoint.py
   

---

Contact & Inquiry

• Email: IgnisLabsReachouts@gmail.com
• Instagram: interestingnights
• Patents: Patents pending.