| --- |
| license: apache-2.0 |
| language: |
| - en |
| tags: |
| - peft |
| - lora |
| - sft |
| - legal |
| - finance |
| - adaption |
| - instruction-tuning |
| base_model: togethercomputer/gpt-oss-20b-bf16 |
| --- |
| |
| # ClauseGuard-Legal-Audit (LoRA Adapter) |
|
|
| ## Model Details |
|
|
| ### Model Description |
| ClauseGuard-Legal-Audit is a specialized Low-Rank Adaptation (LoRA) model fine-tuned for advanced commercial contract analysis. Trained on a premium, sentence-bounded instruction-tuning dataset, this model is engineered to execute precise clause extraction, granular risk tier assessments with legal justifications, and plain-prose business summaries from complex legal documents. |
|
|
| It was developed as part of the Adaption Labs AutoScientist Challenge to benchmark high-tier, agent-trained legal reasoning capabilities. |
|
|
| - **Developed by:** Asad Ullah Dogar |
| - **Model type:** Causal Language Model (LoRA Adapter) |
| - **Language(s) (NLP):** English |
| - **License:** Apache 2.0 |
| - **Finetuned from base model:** `togethercomputer/gpt-oss-20b-bf16` |
| - **Training Platform:** Adaption Labs (AutoScientist Engine) |
|
|
| ### Model Sources |
| - **Dataset Repository:** [adaption-ClauseGuard_Legal_Audit_SFT_v1.0] |
|
|
| ## Uses |
|
|
| ### Direct Use |
| This model is designed for legal professionals, compliance officers, and AI researchers building legal-tech applications. It excels at: |
| * Extracting specific clauses from dense commercial contracts. |
| * Providing granular risk tier assessments (e.g., High, Medium, Low) based on explicit legal text. |
| * Generating step-by-step reasoning traces and legal justifications for its assessments. |
| * Translating complex legal jargon into plain-prose business summaries. |
|
|
| ### Out-of-Scope Use |
| This model is an AI assistant and **does not provide literal or formal legal representation.** It should not be used as a standalone replacement for a licensed attorney. All critical legal assessments must be verified by a human professional. It is not designed for conversational chit-chat outside of the legal/financial domain. |
|
|
| ## Bias, Risks, and Limitations |
| Like all language models, ClauseGuard may occasionally hallucinate or misinterpret highly ambiguous legal phrasing. While it has been fine-tuned using a "Zero Hallucination Constraint" methodology to output "Unavailable" when data is missing, users must still conduct rigorous human-in-the-loop verification for high-stakes audits. |
|
|
| ## How to Get Started with the Model |
|
|
| Use the code below to load the base model and the PEFT LoRA adapters. |
|
|
| ```python |
| from peft import PeftModel |
| from transformers import AutoModelForCausalLM |
| |
| # Load the 20B base model |
| base_model = AutoModelForCausalLM.from_pretrained("togethercomputer/gpt-oss-20b-bf16") |
| |
| # Load the ClauseGuard LoRA adapter |
| model = PeftModel.from_pretrained(base_model, "asadullahdogarr/ClauseGuard-Legal-Audit") |
