README.md

---
tags:
  - text-generation
  - transformers
  - safetensors
  - gguf
  - llama.cpp
  - vllm
  - mlx
  - pytorch
  - onnx
  - llama
  - qwen
  - qwen3_5_text
  - causal-lm
  - scientific-research
  - papers
  - local
  - quantized
  - research-assistant
  - academic-writing
  - latex
  - citations
  - conversational
  - en
  - es
  - zh
  - ja
  - ru
  - fine-tuned
  - finetuned
  - base_model:Qwen/Qwen3.5-4B
  - dataset:Agnuxo/P2PCLAW-Innovative-Benchmark-Agents
  - dataset:Agnuxo/p2pclaw-papers
  - arxiv:2604.19792
  - license:apache-2.0
  - endpoints_compatible
  - region:us
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
---

# CAJAL-4B-P2PCLAW

🧠 **The Research LLM That Fits in Your Pocket**

CAJAL-4B is a 4-billion parameter language model fine-tuned specifically for **scientific paper generation**. Unlike generic chatbots, CAJAL understands academic structure, citation formats, LaTeX, and domain-specific terminology.

Named after **Santiago Ramón y Cajal**, the father of modern neuroscience, this model embodies rigorous, structured thinking applied to scientific writing.

---

## 🚀 Quick Start

### Option 1: HuggingFace Transformers (Python)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Agnuxo/CAJAL-4B-P2PCLAW")
tokenizer = AutoTokenizer.from_pretrained("Agnuxo/CAJAL-4B-P2PCLAW")

prompt = """Write an abstract for a paper on decentralized AI peer review 
using formal verification and IPFS-backed persistence."""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### Option 2: llama.cpp / LM Studio (Local, No Code)

Download the GGUF from [Releases](https://huggingface.co/Agnuxo/CAJAL-4B-P2PCLAW/releases)

Open LM Studio → Load Model → Select GGUF

**System prompt:**
```
You are CAJAL, a research assistant specialized in scientific writing.
Generate well-structured, cited academic content.
Use LaTeX formatting for equations when relevant.
Prefer precise, technical language over vague generalizations.
```

### Option 3: Ollama

```bash
ollama pull agnuxo/cajal-4b-p2pclaw
ollama run agnuxo/cajal-4b-p2pclaw
```

### Option 4: vLLM (Fast Inference Server)

```bash
python -m vllm.entrypoints.openai.api_server \
  --model Agnuxo/CAJAL-4B-P2PCLAW \
  --quantization awq
```

### Option 5: MLX (Apple Silicon)

```python
import mlx_lm

model, tokenizer = mlx_lm.load("Agnuxo/CAJAL-4B-P2PCLAW")
response = mlx_lm.generate(model, tokenizer, prompt="Write a paper abstract...")
```

---

## 📊 What Makes It Different

| Feature | CAJAL-4B | Generic 4B | Why It Matters |
|---------|----------|-----------|---------------|
| **Paper structure** | ✅ Native understanding | ⚠️ Generic chat | Knows IMRAD format |
| **Citations** | ✅ BibTeX, APA, MLA | ❌ Hallucinates | Real citation formats |
| **LaTeX** | ✅ Equations, tables | ❌ No | Research-ready output |
| **Domain terms** | ✅ Physics, CS, Bio | ⚠️ Surface-level | Technical depth |
| **Methodology** | ✅ Detailed procedures | ⚠️ Vague | Reproducible methods |
| **VRAM usage** | ✅ 3.5GB (Q4_K_M) | Similar | Runs on consumer GPUs |
| **Local inference** | ✅ 100% offline | ⚠️ Depends | No API/cloud needed |

---

## 🎯 Benchmarks

| Task | CAJAL-4B | Qwen3.5-4B | Gemma-4B | Phi-4-mini |
|------|----------|-----------|----------|------------|
| Abstract generation | 92/100 | 71/100 | 68/100 | 79/100 |
| Citation accuracy | 88/100 | 52/100 | 48/100 | 61/100 |
| LaTeX correctness | 94/100 | 43/100 | 41/100 | 55/100 |
| Methodology detail | 89/100 | 64/100 | 59/100 | 72/100 |
| Literature review | 85/100 | 69/100 | 67/100 | 74/100 |

Evaluated by [BenchClaw](https://benchclaw.vercel.app) 17-judge tribunal on 50 paper generation tasks.

---

## 💻 Hardware Requirements

| Quantization | File Size | VRAM Required | Speed (RTX 3090) | Speed (M3 Max) |
|-------------|-----------|---------------|-----------------|----------------|
| Q4_K_M | 2.3 GB | 3.5 GB | ~45 tok/s | ~35 tok/s |
| Q5_K_M | 2.7 GB | 4.2 GB | ~42 tok/s | ~32 tok/s |
| Q8_0 | 4.1 GB | 5.0 GB | ~38 tok/s | ~28 tok/s |
| F16 | 8.0 GB | 9.0 GB | ~35 tok/s | ~25 tok/s |

**CPU-only**: Works on any modern CPU. ~5 tok/s on Ryzen 7 5800X.

---

## 🏗️ Training Details

- **Base model**: [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B)
- **Fine-tuning**: LoRA (r=64, alpha=128) on domain-specific corpus
- **Dataset**: 50+ peer-reviewed papers + 200+ technical reports from P2PCLAW ecosystem
- **Context length**: 32K tokens
- **Training hardware**: RTX 3090 24GB
- **License**: Apache 2.0

---

## 🔬 Example Outputs

### Example 1: Abstract Generation

**Prompt**: Write an abstract for a paper on thermodynamic reservoir computing using SHA-256 ASICs.

**Output**:
```
We present a novel architecture for thermodynamic reservoir computing 
that repurposes Bitcoin SHA-256 ASICs as physical substrates. By mapping 
the spin-glass dynamics of the hashing pipeline onto a continuous-time 
reservoir, we achieve energy-efficient computation at the thermodynamic 
limit. Our experimental evaluation on repurposed Antminer S9 hardware 
demonstrates a 2.3x improvement in energy-delay product compared to 
conventional FPGA implementations, with a theoretical bound derived 
from Landauer's principle...
```

### Example 2: Methodology Section

**Prompt**: Write a methodology section for evaluating decentralized AI peer review with 17 judges.

**Output**:
```markdown
### 3.1 Evaluation Protocol

We employ a multi-dimensional tribunal consisting of 17 independent 
judge agents, each specialized in a distinct evaluation criterion:

- **Reasoning Depth** (15%): Logical structure, argument validity, 
  inference chain completeness.
- **Mathematical Rigor** (12%): Proof correctness, notation consistency, 
  theorem applicability.
- **Code Quality** (10%): Reproducibility, documentation, test coverage.
...

Each judge produces a scalar score $s_i \in [0, 100]$ and a confidence 
weight $w_i \in [0, 1]$. The composite score is computed as:

$$S = \frac{\sum_{i=1}^{17} w_i s_i}{\sum_{i=1}^{17} w_i}$$

A paper achieves **Tribunal Pass** if $S \geq 75$ and no individual 
$s_i < 50$ (no veto condition).
```

---

## 🧩 Integration with P2PCLAW Ecosystem

CAJAL is one component of the P2PCLAW distributed research network:

| Component | Role | Link |
|-----------|------|------|
| OpenCLAW-P2P | Core protocol, Lean 4 proofs | [GitHub](https://github.com/Agnuxo1/OpenCLAW-P2P) |
| BenchClaw | 17-judge evaluation | [Web](https://benchclaw.vercel.app) |
| EnigmAgent | Secure credential vault | [GitHub](https://github.com/Agnuxo1/EnigmAgent) |
| AgentBoot | Bare-metal automation | [Web](https://agentboot.pages.dev/) |
| P2PCLAW Main | Research network | [Website](https://www.p2pclaw.com/) |

---

## ⚠️ Limitations

1. **Domain specificity**: Optimized for STEM fields. Less effective for humanities or creative writing.
2. **Hallucination risk**: Like all LLMs, may generate plausible-sounding but incorrect citations. Always verify references.
3. **Language**: Primarily trained on English scientific papers. Spanish, Chinese, Japanese, Russian support is experimental.
4. **Length**: Best for sections up to ~2000 words. Very long papers (>10K words) may lose coherence.
5. **Recency**: Training data cutoff limits knowledge of papers published after training date.

---

## 📚 Citations

If you use CAJAL in research, please cite:

```bibtex
@article{angulo_cajal_2026,
  author = {Angulo de Lafuente, Francisco},
  title = {{CAJAL-4B}: A Research-Specialized Language Model for 
    Decentralized Scientific Writing},
  journal = {arXiv preprint},
  eprint = {2604.19792},
  year = {2026},
  url = {https://arxiv.org/abs/2604.19792}
}
```

---

## 🤝 Contributing

- ⭐ Star the repo: [github.com/Agnuxo1/CAJAL](https://github.com/Agnuxo1/CAJAL)
- 🐛 Report issues: [GitHub Issues](https://github.com/Agnuxo1/CAJAL/issues)
- 💰 Sponsor development: [GitHub Sponsors](https://github.com/sponsors/Agnuxo1)

---

## 📜 License

Apache 2.0 — free for research and commercial use.

---

*Built by [Francisco Angulo de Lafuente](https://www.p2pclaw.com/) · P2PCLAW · Independent Research*

**ORCID**: 0009-0001-1634-7063