narendarcodes/adaption-contract-clause-analyzer-109b

Adaption AutoScientist — Contract Clause Analyzer (109B LoRA)

Powered by Adaptive Data — Adaption Labs

🎯 Core Problem Solved

Given a contract clause or legal fact pattern, this model classifies the clause type across 41 CUAD categories and applies the governing legal rule step-by-step — never fabricating citations or referencing statutes not in the provided text.

Specifically, this model performs:

1. 41-Category CUAD Contract Clause Classification — "Given this contract paragraph, identify whether it contains an IP ownership clause, non-compete restriction, termination provision, indemnification obligation, or any of 41 other CUAD-defined clause types"
2. Hearsay Admissibility Determination — "Given this witness statement and the circumstances of the utterance, determine if the statement qualifies as hearsay under evidence rules, applying each element of the hearsay definition"
3. Tax Statutory Entailment — "Given this IRS tax statute and this taxpayer's fact pattern, determine whether the statute entails (supports) or contradicts the taxpayer's position, evaluating each statutory element against the facts"

Why This Matters

Legal professionals spend hundreds of hours per deal manually reviewing contracts clause-by-clause. Current LLMs hallucinate case citations, fabricate legal rules, and provide ungrounded conclusions. This model is trained to:

• Only reason from the text provided — never reference external cases
• Apply IRAC methodology — Issue → Rule → Application → Conclusion
• Show its legal reasoning in structured <think> traces before reaching a <answer>

Model Details

Field	Value
Trained Model Name	adaption_legal_breach_advisory_sample
Base Model	meta-llama/Llama-4-Scout-17B-16E-Instruct (109B MoE)
Training Method	Supervised Fine-Tuning (SFT) with LoRA
Training Platform	Adaption Labs AutoScientist
Language	English (en)
License	CC-BY-4.0

AutoScientist Platform Integration

The fine-tuning process was managed through the Adaption Labs AutoScientist platform, applying targeted data engineering and training constraints:

• Adaptive Data Pipeline: Enhanced dataset consistency and resolved syntax issues, upgrading the final data quality score from Grade B (9.0) to Grade A (9.8), representing an 8.9% improvement.
• Prompt Deduplication: Sanitized dataset to eliminate duplicate legal fact patterns, preventing the model from overfitting on specific contract types.
• Prompt Rephrase Constraint: Per system guidelines, prompt rephrasing was disabled to maintain statutory definitions and precise contract language.
• Reasoning Traces & Blueprint Constraints: Configured the model to utilize IRAC-style reasoning structures (Issue, Rule, Application, Conclusion). The blueprint enforces:
  1. Declaring the governing legal rule or contract clause category.
  2. Evaluating the fact pattern against the rule elements step-by-step.
  3. Strict grounding of conclusions in the provided context (refusing to cite external statutes or cases).
• Hyperparameter Optimization: Managed the SFT training configuration with LoRA targeting the MoE modules of meta-llama/Llama-4-Scout-17B-16E-Instruct.
• Evaluation: Validated the adapted model against the base model on held-out legal prompts using automated preference scoring.

📊 Dataset

• Dataset: narendarcodes/adaption-contract-clause-analyzer-dataset
• Size: 800 rows
• License: CC-BY-4.0

Source Data & Attribution:

Source	Rows	What It Teaches	License
zenml/cuad-deepseek	~550	41-category contract clause classification with DeepSeek reasoning	Apache-2.0
nguha/legalbench (hearsay)	~125	Hearsay admissibility under evidence rules	CC-BY-4.0
nguha/legalbench (sara_entailment)	~125	Tax statute entailment vs. fact patterns	CC-BY-4.0

Metric	Before	After	Change
Grade	B	A	⬆️
Score	9.0	9.8	+8.9%
Percentile	—	78.4%	—

⚙️ Training Configuration

Hyperparameter	Value
Finetune Job ID	38b574a3-841b-486a-9371-cd1b9ef6f759
Training Experiment ID	0ded6a5c-5597-4c3a-9ad1-3f3e4d6ce5eb
LoRA Rank (r)	64
LoRA Alpha	128
LoRA Dropout	0
Target Modules	q_proj, k_proj, v_proj, o_proj, shared_expert.*, feed_forward.*
Epochs	4
Batch Size	max
Learning Rate	5e-5
LR Scheduler	Cosine (0.5 cycles)
Warmup Ratio	0.05
Weight Decay	0.02
Max Grad Norm	1.0

📈 Evaluation Results

Evaluation Set	Base Model	Adapted Model	Winner
On Your Dataset	19%	81%	✅ Adapted
Across Category (Held-out)	35%	66%	✅ Adapted

How to Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-4-Scout-17B-16E-Instruct")
model = PeftModel.from_pretrained(base_model, "narendarcodes/adaption-contract-clause-analyzer-109b")
tokenizer = AutoTokenizer.from_pretrained("narendarcodes/adaption-contract-clause-analyzer-109b")

Citation

@misc{golla2026contractanalyzer,
  title={Contract Clause Analyzer — CUAD Classification and Legal Rule Application via Grounded Reasoning},
  author={Golla Narendar},
  year={2026},
  note={Trained using Adaption Labs AutoScientist platform. Powered by Adaptive Data.}
}

---

Powered by Adaptive Data — Adaption Labs

Built for the 2026 Adaption AutoScientist Challenge — Legal Category

Framework Versions

• PEFT 0.15.1