---
license: llama3.2
base_model: unsloth/Llama-3.2-1B-bnb-4bit
tags:
- unsloth
- sft
- trl
- lora
- arxiv
- text-generation
- domain-expert
- machine-learning
datasets:
- danivpv/ml-arxiv-instruct
- mlabonne/FineTome-Alpaca-100k
language:
- en
pipeline_tag: text-generation
---

# Llama-ML-Expert-Instruct-1b

An instruction fine-tuned Small Language Model (SLM) specialized in Machine Learning
domain expertise. Fine-tuned from `unsloth/Llama-3.2-1B-bnb-4bit` using LoRA adapters —
efficient enough to run on modest hardware while retaining strong ML-domain reasoning.

Part of the [LLM-ArXiv-Domain-Expert](https://github.com/danivpv/LLM-ArXiv-Domain-Expert)
pipeline, which builds domain-expert LLMs from ArXiv papers end to end: paper parsing →
instruction/preference dataset generation → SFT → (optional) DPO.

## Model Details

| | |
| :--- | :--- |
| **Base model** | [`unsloth/Llama-3.2-1B-bnb-4bit`](https://huggingface.co/unsloth/Llama-3.2-1B-bnb-4bit) (Llama 3.2, 1B) |
| **Fine-tuning method** | LoRA — rank 32, alpha 32, dropout 0.0 |
| **LoRA target modules** | `q_proj, k_proj, v_proj, up_proj, down_proj, o_proj, gate_proj` |
| **Optimizer** | `adamw_8bit` |
| **Prompt format** | Alpaca (`### Instruction: ... ### Response: ...`) |

## Training Data

- [`danivpv/ml-arxiv-instruct`](https://huggingface.co/datasets/danivpv/ml-arxiv-instruct) — instruction-answer pairs synthesized from ArXiv ML papers
- 10,000-sample subset of [`mlabonne/FineTome-Alpaca-100k`](https://huggingface.co/datasets/mlabonne/FineTome-Alpaca-100k) — preserves general instruction-following alongside domain specialization

## How to Use

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="danivpv/Llama-ML-Expert-Instruct-1b",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

prompt = """### Instruction:
{your ML question here}

### Response:
"""
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Limitations

- 1B-parameter base model: strong for domain-specialized short-form Q&A, not a
  general-purpose reasoning or long-context model.
- Trained on synthetically generated instruction data (see
  [`ml-arxiv-instruct`](https://huggingface.co/datasets/danivpv/ml-arxiv-instruct) for
  the generation pipeline) — inherits whatever biases or gaps exist in the source LLM
  used to generate that data.

## Related

- Dataset (SFT): [`danivpv/ml-arxiv-instruct`](https://huggingface.co/datasets/danivpv/ml-arxiv-instruct)
- Dataset (DPO): [`danivpv/ml-arxiv-dpo`](https://huggingface.co/datasets/danivpv/ml-arxiv-dpo)
- Pipeline code: [github.com/danivpv/LLM-ArXiv-Domain-Expert](https://github.com/danivpv/LLM-ArXiv-Domain-Expert)