52 lines
2.5 KiB
Markdown
52 lines
2.5 KiB
Markdown
|
---
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
base_model:
|
|||
|
|
- meta-llama/Llama-3.1-8B-Instruct
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
tags:
|
|||
|
|
- llama
|
|||
|
|
- legal
|
|||
|
|
- marketing
|
|||
|
|
- qlora
|
|||
|
|
- axolotl
|
|||
|
|
---
|
|||
|
|
## Model summary
|
|||
|
|
|
|||
|
|
| | |
|
|||
|
|
|--|--|
|
|||
|
|
| **Type** | Causal LM (merged full weights: base + LoRA) |
|
|||
|
|
| **Base model** | [meta-llama/Meta-Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct) |
|
|||
|
|
| **Task** | Short-form **legal marketing** and **client-facing** copy (website-style tone, practice descriptions, alerts-style prose) |
|
|||
|
|
| **Training** | Supervised fine-tuning (**QLoRA** via [Axolotl](https://github.com/axolotl-ai-cloud/axolotl)); LoRA adapters merged into the base for serving |
|
|||
|
|
| **Language** | English |
|
|||
|
|
| **License** | Use of **Llama** weights is subject to **Meta’s Llama license** and Hugging Face acceptance flow. This adapter/merged artifact is shared under the terms you set on the Hub; the **GitHub project** uses MIT for code/docs—see repo `LICENSE` / `NOTICE`. |
|
|||
|
|
|
|||
|
|
## Intended use
|
|||
|
|
|
|||
|
|
- Drafting or refining **marketing-oriented** legal content (e.g. practice blurbs, client-facing summaries).
|
|||
|
|
- **Not** for legal advice, regulated filings, or high-stakes decisions without human review.
|
|||
|
|
|
|||
|
|
## Training data (high level)
|
|||
|
|
|
|||
|
|
- Data came from **public** law-firm web marketing pages across many large-firm domains, plus an **LLM-assisted curation** step to standardize tone and structure into chat-format SFT pairs.
|
|||
|
|
- **Raw scrapes and full training JSONL are not redistributed** with the GitHub project; statistics and methodology are described in the linked repository.
|
|||
|
|
|
|||
|
|
## Limitations
|
|||
|
|
|
|||
|
|
- **Style and fluency**, not factual grounding: the model can still hallucinate or misstate facts; always verify against sources and counsel.
|
|||
|
|
- Strongest fit for **external-facing, polished** marketing tone; may be less ideal for purely operational or highly technical internal briefs.
|
|||
|
|
- **Bias and safety:** inherits behaviors and limitations of the base Llama 3.1 instruct model; apply usual content policies.
|
|||
|
|
|
|||
|
|
## How to reproduce / cite the project
|
|||
|
|
|
|||
|
|
- GitHub (configs, scripts, evaluation examples): link your public **`fine-tuning-llama-public`** repository when published.
|
|||
|
|
- Base model and Axolotl citations should follow their respective licenses and papers/docs.
|
|||
|
|
|
|||
|
|
## Inference
|
|||
|
|
|
|||
|
|
- Suitable for **vLLM**, **Transformers**, or other Llama-compatible stacks; use the same chat template / tokenizer as **Meta-Llama-3.1-8B-Instruct** unless your serving stack overrides it.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
*This file lives in the GitHub repo as documentation to paste into the Hub; the canonical model page is on Hugging Face.*
|