Downloads · 30 days
0
progressive-cognitive-dream-lora
dexmac/progressive-cognitive-dream-lora
progressive-cognitive-dream-lora is a text generation model from dexmac. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
Modello principale italiano — Qwen2.5-1.5B addestrato con architettura cognitiva progressiva a 4 fasi + SVD Dream Pruning (rank 16→8).
Access
Public
Updated Feb 28, 2026
Repo size
46.4 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt17.5 MB · 38%
At a glance
- Task
- Text Generation
- Library
- peft
- License
- apache-2.0
- Access
- Public
- Created
- Feb 23, 2026
- Updated
- Feb 28, 2026
- SHA
- dee857b7
Base models
Datasets
Model card
From the Hugging Face model README
Architettura Cognitiva Progressiva — Dream-LoRA con SVD Pruning (Italiano)
Modello principale italiano — Qwen2.5-1.5B addestrato con architettura cognitiva progressiva a 4 fasi + SVD Dream Pruning (rank 16→8).
📊 Risultati
| Metrica | Dream-LoRA (questo) | Progressive-LoRA | Flat-LoRA |
|---|---|---|---|
| Accuratezza Esatta | 58.6% ± 2.9 | 37.0% ± 0.5 | 60.6% |
| Number Sense | 60.0% ± 0.8 | 57.7% ± 0.5 | 0.0% |
| Metacognizione | 100.0% | 98.5% | 0.0% |
Il passaggio da magnitude pruning a SVD Dream Pruning ha migliorato significativamente l'accuratezza esatta (+21.6pp) preservando number sense e metacognizione.
🧠 Progressive Cognitive Architecture
A bio-inspired 4-phase training methodology:
| Phase | Name | What happens |
|---|---|---|
| 1 | Foundation | Learn exact arithmetic via LoRA fine-tuning |
| 2 | Consolidation | SVD Dream Pruning (rank 16→8) compresses knowledge into intuition |
| 3 | Delegation | Learn complexity-aware routing: compute internally vs. delegate to tool |
| 4 | Orchestration | Full pipeline: intuit → route → tool → validate |
Guiding Principle: Knowledge doesn't disappear — it collapses into attractors. Intuition is the compressed residue of experience.
🌙 Dream Pruning (Fattorizzazione SVD a Basso Rango)
Invece di azzerare i pesi piccoli, il Dream Pruning usa la decomposizione SVD per ridurre il rango effettivo delle matrici LoRA da 16 a 8. Preserva le direzioni principali ("connessioni logiche") scartando il rumore — analogo al consolidamento della memoria durante il sonno.
🔧 Configurazione
| Parametro | Valore |
|---|---|
| Modello Base | Qwen/Qwen2.5-1.5B |
| LoRA Rank | 16 (→ 8 dopo SVD) |
| LoRA Alpha | 32 |
| Target LoRA | q_proj, k_proj, v_proj, o_proj |
| Tipo Pruning | SVD Low-Rank Factorization |
| Lingua Dati | Italiano |
🚀 Uso Rapido
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2.5-1.5B", device_map="auto", torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B")
model = PeftModel.from_pretrained(
base_model,
"dexmac/progressive-cognitive-dream-lora",
subfolder="lora_adapters"
)
messages = [{"role": "user", "content": "Risolvi: 342 * 67"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
🔗 Modelli Correlati
- Progressive-LoRA (IT) — Primo prototipo con magnitude pruning
- Flat-LoRA (IT) — Controllo senza fasi
- 1.5B Dream (EN) — Miglior modello (inglese, composite 87.6)
- GitHub — Codice sorgente completo
📝 Citation
@software{progressive_cognitive_2026,
author = {Dex Mac},
title = {Progressive Cognitive Architecture for LLMs},
year = {2026},
url = {https://github.com/dexmac221/progressive-cognitive},
version = {1.0.0}
}
📄 License
Apache 2.0
- Task
- Text Generation
- Library
- peft
- License
- apache-2.0
- Languages
- it
- Created
- Feb 23, 2026
- Updated
- Feb 28, 2026