Fine-tuning data quality is signed. Sparse curation lobotomizes; dense curation teaches the model how to access and articulate its own pre-trained knowledge.

Papers

#	Title	Role	Status
1	The Entropy of Recursion	Theoretical foundation — defines PEST, PPP, NEST, Model Collapse, Context Blindness	Published
2	The Density Imperative	Empirical validation — the ablation study with live results	This experiment
3	Cognitive Nutrition: Architecture of a Provenance-Embedded Data Refinery	Systems/dataset paper — the factory that produces the cure	Published

Author: Tad MacPherson, Metavolve Labs, Inc. Base Model: Llama 3.2 11B Vision-Instruct Dataset: Alexandria Aeternum Genesis (10,090 artworks from The Metropolitan Museum of Art) Interactive Explorer: iaeternum.ai/evaluation — browse all 100 evaluation images × 3 models × 8 evaluation tracks

The Experiment

A controlled ablation study proving that fine-tuning data quality is not a scalar quantity but a signed intervention. We fine-tuned Llama 3.2 11B Vision-Instruct on 9,081 cultural heritage images under three conditions — holding images, architecture, and hyperparameters constant. The sole variable is metadata density.

Condition	Label	Tokens/Image	Description
Base	No fine-tuning	N/A	Llama 3.2 11B Vision-Instruct, unmodified
Group A	Sparse captions	~50–100	Title, creator, date, medium (simulates web-scraped alt-text)
Group B	Dense NEST	~2,000–4,000	111-field schema: composition, technique, cultural context, affect, symbolism, narrative

Results

The Three-Way Gradient

Metric	Base	Group A (Sparse)	Group B (Dense)	A vs Base	B vs Base
VULCA Overall	0.529	0.329	0.559	−37.8%	+5.6%
CogBench Overall	0.384	0.175	0.418	−54.4%	+8.9%
Semantic Coverage	0.099	0.028	0.257	−71.7%	+160.3%
Hallucination Rate	0.010	0.043	0.013	+330%	+33%
Counterfactual Rejection	0.933	0.455	0.765	−51.3%	−18.0%
Avg Tokens	160.7	37.9	136.0	−76.4%	−15.4%

CogBench Cognitive Depth (Track 4)

All 8 dimensions statistically significant: Friedman test, all p < .001, n = 100.

Dimension	Base	Group A	Group B	χ²
Overall	0.378	0.174	0.415	157.22
Narrative	0.280	0.175	0.413	151.21
Mental State	0.392	0.183	0.492	174.24
Focal Point	0.454	0.165	0.567	189.23
Atmosphere	0.227	0.162	0.309	82.99
Cultural Knowledge	0.412	0.192	0.378	130.03
Color Function	0.472	0.165	0.362	152.83
Causal Depth	0.407	0.178	0.388	136.22

Semantic Density Coverage

Field	Base	Group A	Group B	B vs Base
Composition	0.132	0.022	0.317	+140.5%
Color Harmony	0.130	0.015	0.341	+162.7%
Technique	0.104	0.021	0.221	+112.4%
Emotional	0.067	0.040	0.256	+282.1%
Symbolic	0.098	0.033	0.224	+128.9%
Narrative	0.062	0.034	0.183	+196.3%
Overall	0.099	0.028	0.257	+160.3%

GLMM Verbosity Control

Parameter	β	SE	z	p
Intercept	0.247	0.007	33.94	<.001
Group A (vs Base)	−0.106	0.006	−16.81	<.001
Group B (vs Base)	0.003	0.004	0.69	.488 (ns)
Token count	0.000075	0.000004	19.51	<.001

Key insight: Group A's destruction is real and independent of verbosity (β = −0.106, p < .001). Group B's advantage is mediated by the structured analytical patterns the model acquires — token count is a mediator, not a confounder.

Training Details

Group A (Sparse)

Records: 9,081
Steps: 1,135 (~1 epoch)
Final loss: 0.865
Training time: 34.9 hours
GPU: NVIDIA L4

Group B (Dense NEST)

Records: 9,081
Steps: 1,135 (~1 epoch)
Final loss: 1.164
Training time: 38.3 hours
GPU: NVIDIA L4

Shared Hyperparameters

Parameter	Value
Base model	`unsloth/Llama-3.2-11B-Vision-Instruct`
Quantization	4-bit (Unsloth)
LoRA rank	32
LoRA alpha	32
Target modules	q, k, v, o, gate, up, down
Vision encoder	Frozen (no gradient updates)
Effective batch size	8 (batch 1 × gradient accumulation 8)
Learning rate	2e-4
Optimizer	AdamW (8-bit)
Precision	bfloat16
Image resolution	560px max (1 vision tile)

Evaluation Framework

Four tracks, applied to 100 held-out images (909 reserve images available for community replication):

Track 1 — VULCA-Bench (L1–L5): Visual perception, technical knowledge, symbolic understanding, historical context, philosophical depth
Track 2 — POPE-Style Probes: Binary object presence questions (positive + adversarial + random)
Track 3 — Response Quality: Token density, refusal rate
Track 4 — Structural Generalization: CogBench cognitive depth (8 dimensions × 100 images), counterfactual resistance, semantic density alignment (6 NEST fields)

Total: 4,400+ question–answer pairs across three model conditions.

Repository Contents

├── README.md                           # This file
├── group_a_adapter/                    # LoRA adapter — Group A (sparse captions)
│   ├── adapter_config.json
│   ├── adapter_model.safetensors       # ~450MB
│   ├── tokenizer.json
│   ├── tokenizer_config.json
│   └── ...
├── group_b_adapter/                    # LoRA adapter — Group B (dense NEST)
│   ├── adapter_config.json
│   ├── adapter_model.safetensors       # ~450MB
│   ├── tokenizer.json
│   ├── tokenizer_config.json
│   └── ...
├── evaluation/                         # Evaluation results from GCS
│   ├── cn_experiment_results.json      # T1-3 results (early — see results/ for merged)
│   ├── track4_results.json             # T4 per-model results
│   ├── track4_results_merged.json      # T4 all 3 conditions merged
│   └── track4_raw_responses.json       # Raw T4 model outputs
├── results/                            # Complete results package
│   ├── cn_experiment_results_t13.json  # Merged T1-3 (Base + A + B)
│   ├── track4_results_merged.json      # Merged T4 (Base + A + B)
│   ├── track4_raw_responses_base_and_a.json  # Raw Q&A: Base + Group A
│   ├── track4_raw_responses_groupb.json      # Raw Q&A: Group B
│   ├── statistical_analysis_results.json     # Friedman, Wilcoxon, GLMM
│   ├── ppp_survival_test_results.json        # Perceptual hash survival
│   ├── STATISTICAL_ANALYSIS.md               # Human-readable stats
│   ├── CN_EXPERIMENT_RESULTS_COMPLETE.md     # Full results narrative
│   └── training_metrics/
│       ├── group_a_training.csv
│       ├── group_b_training.csv
│       └── training_metrics.json
├── scripts/                            # Full reproducibility scripts
│   ├── train_cn_experiment.py          # Training script
│   ├── prepare_data.py                 # Data preparation (Group A/B split)
│   ├── evaluate_cn.py                  # Tracks 1-3 evaluation
│   ├── evaluate_track4.py              # Track 4 evaluation
│   ├── statistical_analysis.py         # Statistical analysis (Friedman, GLMM)
│   └── cn_experiment_startup.sh        # VM orchestrator
├── eval_explorer.html                  # Interactive comparison tool (3.3MB standalone)
├── group_a_stats.json                  # Training stats
└── group_b_stats.json                  # Training stats

Interactive Eval Explorer

The eval_explorer.html file is a self-contained, 3.3MB interactive tool that lets you browse:

All 100 evaluation images
Side-by-side responses from Base, Group A, and Group B
All 8 CogBench dimensions + counterfactual + semantic density tracks
Per-image scores and comparisons

Open it in any browser — no server required. Also hosted at iaeternum.ai/evaluation.

Using the Adapters

python

from unsloth import FastVisionModel
from peft import PeftModel

# Load base model
model, tokenizer = FastVisionModel.from_pretrained(
    "unsloth/Llama-3.2-11B-Vision-Instruct",
    load_in_4bit=True,
)

# Apply Group B (NEST-enriched) adapter
model = PeftModel.from_pretrained(
    model,
    "Metavolve-Labs/cognitive-nutrition-experiment",
    subfolder="group_b_adapter"
)
FastVisionModel.for_inference(model)

Reproducing the Experiment

bash

# 1. Get the dataset
pip install datasets
from datasets import load_dataset
ds = load_dataset("Metavolve-Labs/alexandria-aeternum-genesis")

# 2. Prepare training data (creates Group A sparse + Group B NEST splits)
python scripts/prepare_data.py --input alexandria_aeternum_full_v2.jsonl

# 3. Train (requires NVIDIA L4/A100, 24GB+ VRAM)
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
python scripts/train_cn_experiment.py --group A
python scripts/train_cn_experiment.py --group B

# 4. Evaluate
python scripts/evaluate_cn.py           # Tracks 1-3
python scripts/evaluate_track4.py       # Track 4 (CogBench, counterfactual, semantic density)
python scripts/statistical_analysis.py  # Statistical tests

The Key Insight

Dense structured training data teaches the model how to think about a domain, not merely what to say about specific examples. Group B produces superior analyses of images the model has never seen during training. The knowledge about art history, composition, and technique exists in both models' pre-trained weights. Only Group B has learned to access and organize that knowledge into structured analytical responses. This is not knowledge injection — it is cognitive scaffolding.

Citation

bibtex

@article{macpherson2026density,
  title={The Density Imperative: How Semantic Curation Depth Determines Vision-Language Model Capability},
  author={MacPherson, Tad},
  year={2026},
  institution={Metavolve Labs, Inc.}
}

@article{macpherson2025entropy,
  title={The Entropy of Recursion: A Strategic Framework for Provenance-Verified Data and the Preservation of Signal Integrity in Generative Systems},
  author={MacPherson, Tad},
  year={2025},
  institution={Metavolve Labs, Inc.}
}

@article{macpherson2026architecture,
  title={Cognitive Nutrition: Architecture of a Provenance-Embedded Data Refinery},
  author={MacPherson, Tad},
  year={2026},
  institution={Metavolve Labs, Inc.}
}

Patent

U.S. Provisional Patent Application No. 63/983,304 — "Systems and Methods for Provenance-Embedded Semantic Transport, Multi-Agent Data Enrichment, and Perceptual Compute Offloading in Autonomous Systems"

License

Apache 2.0. The LoRA adapters, evaluation scripts, and results are freely available for research and commercial use. The underlying artworks are public domain (Met Museum Open Access).

Built by Metavolve Labs — Cognitive Nutrition for Foundation Models. "The factory is open."