Pluto-Genesis-0.6B/README.md

---
license: apache-2.0
language:
  - en
tags:
  - pluto-genesis
  - qwen3
  - sub-1b
  - qlora
  - instruction-tuning
  - reasoning
  - mathematics
  - coding
  - fine-tuned
  - conversational
  - text-generation-inference
base_model: Qwen/Qwen3-0.6B
pipeline_tag: text-generation
library_name: transformers
datasets:
  - teknium/OpenHermes-2.5
  - microsoft/orca-math-word-problems-200k
  - m-a-p/CodeFeedback-Filtered-Instruction
model-index:
  - name: Pluto-Genesis-0.6B
    results: []
---

<div align="center">

# 🪐 Pluto-Genesis-0.6B

<p align="center">
  <img src="./banner.png" width="100%">
</p>

**An instruction-tuned sub-1B language model fine-tuned on 80K curated samples**

[![HuggingFace](https://img.shields.io/badge/🤗-HuggingFace-yellow)](https://huggingface.co/Siddh07ETH/Pluto-Genesis-0.6B)
[![License](https://img.shields.io/badge/License-Apache%202.0-blue)](https://opensource.org/licenses/Apache-2.0)
[![Model Size](https://img.shields.io/badge/Parameters-596M-green)]()
[![Base Model](https://img.shields.io/badge/Base-Qwen3--0.6B-orange)](https://huggingface.co/Qwen/Qwen3-0.6B)
[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21368749.svg)](https://doi.org/10.5281/zenodo.21368749)

</div>

---

## Model Description

**Pluto-Genesis-0.6B** is a fine-tuned instruction-following language model built on top of [Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B). It was trained using **QLoRA** (Quantized Low-Rank Adaptation) on a curated mixture of 80,000 high-quality instruction-response pairs spanning general reasoning, mathematical problem solving, and code generation.

This model is part of the **Pluto AI** research project by Siddharth N.R., exploring efficient fine-tuning of sub-1B language models on consumer-grade hardware.

> **Research Goal:** Demonstrate that a sub-1B model fine-tuned on carefully curated data can achieve competitive performance on reasoning, math, and coding benchmarks while remaining deployable on consumer hardware.

---

## Training Details

| Property | Value |
|----------|-------|
| **Base Model** | Qwen/Qwen3-0.6B |
| **Parameters** | 596 Million (596,049,920) |
| **Method** | QLoRA (4-bit NF4 + LoRA) |
| **LoRA Rank** | r=64, α=128 |
| **LoRA Target Modules** | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| **Training Steps** | 2,475 |
| **Final Training Loss** | 0.2741 |
| **Precision** | FP16 |
| **Optimizer** | Paged AdamW 8-bit |
| **Learning Rate** | 2e-4 (cosine schedule) |
| **Effective Batch Size** | 32 (2 × 16 grad accum) |
| **Sequence Length** | 1024 tokens |
| **Hardware** | Tesla T4 (16 GB) |
| **Framework** | Transformers + PEFT + TRL |

---

## Training Data

| Domain | Dataset | Samples | Skills Targeted |
|--------|---------|---------|----------------|
| 🧠 General Reasoning | [OpenHermes-2.5](https://huggingface.co/datasets/teknium/OpenHermes-2.5) | 30,000 | Instruction following, reasoning |
| 🔢 Mathematics | [Orca-Math-200K](https://huggingface.co/datasets/microsoft/orca-math-word-problems-200k) | 30,000 | Word problems, step-by-step math |
| 💻 Code | [CodeFeedback-Filtered](https://huggingface.co/datasets/m-a-p/CodeFeedback-Filtered-Instruction) | 20,000 | Code generation, debugging |
| **Total** | | **80,000** | |

---

## Benchmarks

> 📊 Benchmarks will be added shortly. The model is currently being evaluated on ARC-Challenge, HellaSwag, MMLU, GSM8K, and TruthfulQA using lm-evaluation-harness v0.4.4.

<p align="center">
  <img src="./detailed_benchmark.png" width="100%">
</p>

<p align="center">
  <img src="./benchmark-comparison.png" width="100%">
</p>

---

## Usage

## GGUF Quantizations

GGUF versions for llama.cpp, Ollama and LM Studio are available here:

> https://huggingface.co/Siddh07ETH/Pluto-Genesis-0.6B-GGUF

### Basic Inference

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

model = AutoModelForCausalLM.from_pretrained(
    "Siddh07ETH/Pluto-Genesis-0.6B",
    torch_dtype=torch.float16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Siddh07ETH/Pluto-Genesis-0.6B")

messages = [{"role": "user", "content": "Explain what a neural network is."}]

text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.3,
        do_sample=True,
        top_p=0.9,
        repetition_penalty=1.1,
    )

response = tokenizer.decode(
    output[0][inputs.input_ids.shape[1]:],
    skip_special_tokens=True
)
print(response)
```

### Low Memory Inference (4-bit)

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

quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)
model = AutoModelForCausalLM.from_pretrained(
    "Siddh07ETH/Pluto-Genesis-0.6B",
    quantization_config=quant_config,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Siddh07ETH/Pluto-Genesis-0.6B")
```

---

## Recommended Generation Settings

| Setting | Value | Reason |
|---------|-------|--------|
| `temperature` | 0.3 | Conservative — reduces hallucinations |
| `top_p` | 0.9 | Focused vocabulary |
| `repetition_penalty` | 1.1 | Prevents rambling |
| `max_new_tokens` | 200–512 | Keeps answers concise |
| `do_sample` | True | Required when temperature < 1.0 |

---

## Limitations

- **Model size:** At 596M parameters this model will hallucinate on topics outside its training distribution. Always verify factual claims.
- **Context length:** Trained on sequences up to 1024 tokens. Performance may degrade on longer contexts.
- **Knowledge cutoff:** The model does not have access to real-time information.
- **Research only:** Not intended for production deployment without further evaluation and safety testing.

---

## Author

**Siddharth N.R.**
[HuggingFace](https://huggingface.co/Siddh07ETH)

---

## 📄 Research Paper

The complete research paper describing the Pluto-Genesis training pipeline, benchmark evaluation, checkpoint recovery system, and engineering methodology is available on Zenodo.

**Paper**
> Pluto-Genesis: Reliable QLoRA Fine-Tuning on Ephemeral Compute with Cross-Session Checkpoint Recovery for Sub-1B Language Models

**DOI**
https://doi.org/10.5281/zenodo.21368749

This paper includes:

- Training methodology
- QLoRA configuration
- Cross-session checkpoint recovery
- Benchmark evaluation
- Deployment workflow
- Reproducibility details

## Citation

```bibtex
@misc{plutogenesis2026,
  author    = {Siddharth N.R.},
  title     = {Pluto-Genesis-0.6B: An Instruction-Tuned Sub-1B Language Model},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/Siddh07ETH/Pluto-Genesis-0.6B}
}
```

---

## License

Apache 2.0 — see [LICENSE](https://opensource.org/licenses/Apache-2.0).
Base model [Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) is also Apache 2.0.