You can deploy this model and start earning money today!

narcolepticchicken/
contract-drafting-assistant-v2

Name: contract-drafting-assistant-v2
Availability: InStock

Views

No views yet

Contract Drafting Assistant — Complete System

A RAG-augmented LLM pipeline for drafting full legal contracts from deal context, party position, and business constraints.

System Architecture

Deal Context → Playbook Rules → Dense Retriever (37K clauses) → SaulLM-7B (4-bit) → NLI Verifier → Final Contract
     ↑                              ↑                                      ↑                 ↑
Party Position             Clause Dependency Graph            Precedent Clauses      Per-Clause NLI

Supported Contract Types

SaaS Agreement
Master Services Agreement (MSA)
Non-Disclosure Agreement (NDA)
Data Processing Agreement (DPA)
Statement of Work (SOW)
Vendor Agreement
Consulting Agreement
IP Assignment
Employment/Contractor Agreement

Key Components

Playbook Rules (`playbook.py`)

Encodes expert knowledge as structured rules:

Required clauses per contract type (9 contract types, ~12 clauses each)
Fallback positions per clause × party position (pro_company, balanced, pro_counterparty)
Risk flags with severity levels (HIGH/MEDIUM)
Drafting checklist per contract type (~8-10 items each)

Dense Retriever (`dense_retriever.py`)

Hybrid BM25 + cosine similarity over 37,508 real contract clauses from EDGAR/ALM. Used to provide precedent examples when drafting each clause.

Clause Dependency Graph (`clause_graph.py`)

14 directed edges modeling how clauses interact:

CONSTRAINS: liability cap limits indemnification
REQUIRES: termination needs defined breach conditions
TRIGGERS: data protection activates subprocessor rules
MATCHES: governing law must match arbitration seat

NLI Verifier

Per-clause entailment checking against 100+ legal hypotheses using roberta-large-mnli. Future: swap to contract-specific NLI (contract-nli-v1).

Fine-tuned Model (`contract-drafter-v1`)

Saul-7B-Instruct-v1 fine-tuned on 126 contracting examples across 9 types. See narcolepticchicken/contract-drafter-v1.

Evaluation Results

Evaluated on 7 gold-standard drafting tasks with rubric scoring across 8 metrics.

Version	Model	Overall	NLI	Semantic Recall
V11	Base SaulLM-7B	0.646	0.553	0.167
V12 (running)	Fine-tuned	TBD	TBD	TBD

See contract-drafts-v1 for full results.

Known Issues (V11)

key_term_recall_semantic = 0.167: NLI truncates full contract to 2000 chars
roberta-large-mnli: General-purpose NLI misses legal entailments (~55% precision)
Semantic recall: Only checks first ~4 clauses, misses terms in later clauses

Planned Improvements

Per-clause key term matching (+0.086 overall)
Swap to contract-nli-v1 (+0.022 overall)
Lower eval temperature to 0.1 for cleaner measurements
Seed deal memory for counterparty history

Hub Assets

Repo	Type	Description
contract-drafter-v1	Model	Fine-tuned drafting model (3.8GB)
contract-clause-index-v1	Dataset	37.5K clauses + embeddings
contract-nli-v1	Model	Legal-domain NLI
contract-drafts-v1	Dataset	Eval results (v8-v12)
contract-drafting-assistant	Model	Playbook + fallback positions
contract-drafting-assistant	Space	Interactive Gradio UI

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
    "narcolepticchicken/contract-drafter-v1",
    device_map="auto",
    load_in_4bit=True,
)
tokenizer = AutoTokenizer.from_pretrained("narcolepticchicken/contract-drafter-v1")
# See run_v12.py for the full drafting pipeline

narcolepticchicken/contract-drafting-assistant-v2