narcolepticchicken/contract-drafter-v1

Contract Drafter v1

Fine-tuned Equall/Saul-7B-Instruct-v1 for legal contract clause drafting.

Training

• Method: SFT + LoRA (r=16, alpha=32)
• Examples: 126 across 9 contract types (SaaS, MSA, NDA, DPA, SOW, Vendor, Consulting, IP Assignment, Employment)
• Epochs: 3
• Learning rate: 2e-4 with cosine schedule
• Final loss: 0.568 (down from 2.608)
• Token accuracy: 96.9%

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
    "narcolepticchicken/contract-drafter-v1",
    device_map="auto",
    torch_dtype="auto",
)
tokenizer = AutoTokenizer.from_pretrained("narcolepticchicken/contract-drafter-v1")

prompt = """<s>[INST] You are an expert contract drafter.

Draft a limitation_of_liability clause for a SaaS agreement.
Deal: Enterprise SaaS platform, $200K ACV, SOC 2 Type II.
Constraints: annual billing, 99.9% uptime SLA.
Law: Delaware

Draft ONLY the clause text. [/INST]"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=600, temperature=0.3)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Performance

Evaluated against base model on 7 contract types. See contract-drafts-v1 for full results.

Files

• model.safetensors — merged weights (3.8GB, 4-bit)
• lora/ — LoRA adapter for future fine-tuning
• tokenizer.json — SaulLM tokenizer (Mistral-based)

Related

• contract-drafting-assistant Space — interactive drafting UI
• contract-clause-index-v1 — 37,508 real contract clauses
• contract-nli-v1 — legal-domain NLI model