metadata
language: en
license: apache-2.0
library_name: peft
base_model: Qwen/Qwen2.5-3B
tags:
- lora
- peft
- baseline
- flat-training
- math
- arithmetic
- control-group
datasets:
- custom
pipeline_tag: text-generation
model-index:
- name: progressive-cognitive-qwen3b-baseline-lora
results:
- task:
type: text-generation
name: Cognitive Arithmetic
metrics:
- type: exact_accuracy
value: 60.4
name: Exact Accuracy (%)
- type: composite_score
value: 78.5
name: Composite Cognitive Score
Progressive Cognitive Architecture — 3B Flat LoRA (English, Control)
Control model — Qwen2.5-3B fine-tuned with all training data in a single pass (no phases, no pruning). Serves as the 3B baseline for evaluating progressive training.
📊 Results
| Metric | Score |
|---|---|
| Composite Score | 78.5 |
| Exact Accuracy | 60.4% ± 7.5 |
| Adversarial Robustness | 84.7% ± 1.2 |
| Delegation Accuracy | 100.0% ± 0.0 |
| Delegation Rate | 58.7% ± 4.6 |
| Magnitude Sense (OoM±1) | 84.0% ± 4.0 |
| Catastrophic Errors | 0.0% ± 0.0 |
Results: mean ± std over 3 seeds (42, 43, 44), 50 samples × 5 dimensions per seed.
⚖️ Comparison: Flat vs Dream at 3B
| Metric | 3B Flat (this) | 3B Dream | Delta |
|---|---|---|---|
| Composite | 78.5 | 66.0 | -12.5 |
| Adversarial | 84.7% | 34.0% | -50.7pp |
| Catastrophic | 0.0% | 41.3% | +41.3pp |
At 3B scale, flat training outperforms progressive Dream training — the inverse of the 1.5B result. This supports the hypothesis that SVD compression (rank 16→8) creates adapters too weak relative to the larger base model's weight space.
🔧 Training Configuration
| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen2.5-3B |
| LoRA Rank | 16 |
| LoRA Alpha | 32 |
| LoRA Targets | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Dropout | 0.05 |
| Training Data | ~6,000 English arithmetic examples (all mixed in one pass) |
| Hardware | NVIDIA T4 16GB |
🚀 Quick Start
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2.5-3B", device_map="auto", torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B")
model = PeftModel.from_pretrained(
base_model,
"dexmac/progressive-cognitive-qwen3b-baseline-lora"
)
messages = [{"role": "user", "content": "Calculate: 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))
🔗 Related Models
- 1.5B Dream LoRA — Best overall model
- 3B Dream LoRA — Progressive training on 3B
- 1.5B Flat LoRA — 1.5B control
- Results Dataset — Raw evaluation data
- GitHub — Full source code
📝 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