README.md

---
language:
- en
license: llama3.2
base_model: meta-llama/Llama-3.2-1B
tags:
- llama
- continued-pretraining
- sft
- lora
- 1b
- math
- code
- education
- small-llm
datasets:
- HuggingFaceFW/fineweb-edu
- open-web-math/open-web-math
- bigcode/starcoderdata
- HuggingFaceTB/cosmopedia
- teknium/OpenHermes-2.5
- meta-math/MetaMathQA
- sahil2801/CodeAlpaca-20k
---

# Kybalion-1B

**Kybalion-1B** is a 1B-parameter language model built on top of [Llama 3.2 1B](https://huggingface.co/meta-llama/Llama-3.2-1B) through a full **Continued Pre-Training (CPT) → Supervised Fine-Tuning (SFT)** pipeline, trained entirely on Google Colab A100.

> **Why "Kybalion"?**
> The model was originally developed under the internal codename *Prometheus-1B*, but was renamed to *Kybalion-1B* before public release to avoid confusion with an existing model of the same name on HuggingFace. *Kybalion* refers to the ancient hermetic text symbolizing hidden knowledge — fitting for a model focused on education, mathematics, science, and code.

---

## 🏆 Key Highlights

- **Beats Llama-3.2-1B-Instruct** on HellaSwag (63.8% vs 61.1%) and ties on WinoGrande (62.4%)
- **4.5× GSM8K improvement** over TinyLlama-1.1B (10.8% vs 2.4%) — math pretraining works
- Outperforms TinyLlama-1.1B on **all 6 benchmarks**
- Trained by a single undergraduate student on consumer cloud hardware

---

## 🔬 Key Contributions

- Demonstrates that domain-balanced continued pretraining on curated multi-domain data (education, math, code, science) yields consistent improvements across commonsense reasoning benchmarks in 1B-scale models
- Suggests that multi-step mathematical reasoning remains a fundamental bottleneck for 1B-scale models, even when combining math-focused pretraining (OpenWebMath) with instruction tuning (MetaMathQA)
- Provides a fully reproducible, compute-efficient training recipe (CPT → LoRA SFT) built and executed **by a single undergraduate student in under one week**, demonstrating that meaningful LLM research is achievable without institutional resources or large teams

---

## 📊 Benchmark Results

All scores measured with [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) under **identical conditions** (same prompts, same few-shot settings, same hardware).

| Benchmark | TinyLlama-1.1B | Llama-3.2-1B-Instruct | **Kybalion-1B** |
|-----------|:--------------:|:---------------------:|:---------------:|
| MMLU | 25.0% | 46.1% | **32.0%** |
| ARC-C | 37.2% | 41.5% | **37.6%** |
| GSM8K | 2.4% | 33.5% | **10.8%** |
| HellaSwag | 61.2% | 61.1% | **63.8%** 🏆 |
| WinoGrande | 61.8% | 62.4% | **62.4%** 🏆 |
| TruthfulQA | 37.4% | 43.3% | **40.0%** |

> 🏆 = outperforms Llama-3.2-1B-Instruct
> All evaluations run with `lm_eval.simple_evaluate()`, bfloat16, batch_size=8, A100 GPU.

---

## 🔧 Training Pipeline

### Phase 1: Continued Pre-Training (CPT)

Fine-tuned the base weights of `meta-llama/Llama-3.2-1B` on ~3.5B tokens of curated multi-domain data.

| Domain | Dataset | Ratio | Purpose |
|--------|---------|-------|---------|
| Education | FineWeb-Edu (score ≥ 3.0) | 35% | General knowledge & reasoning |
| Mathematics | OpenWebMath | 20% | Mathematical reasoning |
| Code | StarCoderData (Python) | 15% | Code generation |
| Textbook | Cosmopedia web_samples_v2 | 15% | Structured knowledge |
| Science | Cosmopedia stanford | 10% | Scientific reasoning |
| Story | Cosmopedia stories | 5% | Language fluency |

**Training config:**
- Hardware: Google Colab A100 80GB
- Optimizer: AdamW, LR = 2e-5, Cosine decay, Warmup = 1000 steps
- Precision: BF16
- Effective batch size: 32 (4 × 8 grad accum)
- Sequence length: 2048 (packed)
- Framework: HuggingFace `transformers.Trainer` (no Unsloth)

### Phase 2: Supervised Fine-Tuning (SFT)

Applied LoRA adapters to teach instruction-following, then merged into base weights.

| Dataset | Size | Purpose |
|---------|------|---------|
| OpenHermes 2.5 | 100K | General instruction following |
| MetaMathQA | 50K | Mathematical reasoning (GSM8K boost) |
| CodeAlpaca | 20K | Code generation |

**SFT config:**
- Method: LoRA (r=64, α=128, dropout=0.05)
- Target modules: q/k/v/o/gate/up/down proj (all linear layers)
- LR = 1e-4, Epochs = 3, Cosine decay
- Merged with `PeftModel.merge_and_unload()` for standalone deployment

---

## 💻 Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("devwoo/Kybalion-1B")
model = AutoModelForCausalLM.from_pretrained(
    "devwoo/Kybalion-1B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

def chat(user_message, system="You are a helpful and knowledgeable AI assistant."):
    prompt = (
        f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n"
        f"{system}<|eot_id|>"
        f"<|start_header_id|>user<|end_header_id|>\n\n"
        f"{user_message}<|eot_id|>"
        f"<|start_header_id|>assistant<|end_header_id|>\n\n"
    )
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.7,
            top_p=0.9,
            do_sample=True,
            eos_token_id=tokenizer.convert_tokens_to_ids("<|eot_id|>"),
        )
    return tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)

print(chat("Explain the Pythagorean theorem and give an example."))
print(chat("Write a Python function to check if a number is prime."))
```

---

## 📦 GGUF Version

A quantized **GGUF q4_k_m** version is available at [devwoo/Kybalion-1B-GGUF](https://huggingface.co/devwoo/Kybalion-1B-GGUF) for CPU/mobile inference with [llama.cpp](https://github.com/ggerganov/llama.cpp) or [Ollama](https://ollama.com).

```bash
# With llama.cpp
./llama-cli -m Kybalion-1B-q4_k_m.gguf -p "Explain quantum computing." -n 256
```

---

## ⚠️ Limitations

- 1B parameters — smaller than most production models; may struggle with complex multi-step reasoning
- Not RLHF-aligned; may occasionally produce unhelpful or inconsistent responses
- English-only training data
- GSM8K score (10.8%) reflects room for improvement in math reasoning compared to larger models

---

## 📄 License

This model is derived from `meta-llama/Llama-3.2-1B` and follows the [Llama 3.2 Community License](https://ai.meta.com/llama/license/).
Training datasets are used under their respective open licenses.