StableLex

StableLex is a PEFT/LoRA adapter based on Qwen/Qwen3.5-9B. It is not a standalone language model: it must be loaded together with the base model.

The adapter is intended for controlled stablecoin and virtual-asset legal/compliance analysis workflows where responses are expected to stay grounded in supplied materials, use explicit citations, distinguish source-backed conclusions from cautious inference, and refuse or qualify answers when the provided materials are insufficient.

Repository Contents

• adapter_model.safetensors: LoRA adapter weights.
• adapter_config.json: PEFT adapter configuration.
• formal_eval_report.md: copied formal evaluation report for this adapter release.
• README.md: this model card.
• LICENSE: Apache License 2.0 for this adapter repository.

Model Details

Item	Value
Model name	StableLex
Repository	Karitasu/StableLex
Hub URL	https://huggingface.co/Karitasu/StableLex
Base model	Qwen/Qwen3.5-9B
Adapter type	LoRA via PEFT
Task type	CAUSAL_LM
PEFT version	0.19.1
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
License	Apache License 2.0 for this adapter repository

Intended Use

This adapter is designed for domain-specific answer generation in stablecoin, virtual-asset, and compliance-oriented research settings. Suitable uses include:

• material-grounded analysis of stablecoin and virtual-asset regulatory documents;
• comparison of legal, academic, policy, judicial, and industry materials;
• drafting structured compliance or risk-analysis responses based on supplied source passages;
• identifying when the supplied materials are insufficient to support a legal, regulatory, or factual conclusion;
• citation-sensitive workflows where each substantive claim should be traceable to provided source chunks.

The adapter is best used in a retrieval-augmented or otherwise source-controlled environment. It should receive the relevant source text, source identifiers, and explicit instructions about citation format, material boundaries, and refusal behavior.

Out-of-Scope Use

This adapter is not a substitute for qualified legal, financial, compliance, or professional advice. It should not be used as an autonomous decision-maker for high-stakes matters, licensing analysis, enforcement risk assessment, investment decisions, or client-facing legal opinions without human review.

The released evaluation supports the adapter's behavior under the tested prompt, citation, and material-boundary setup only. Performance outside that setup has not been established.

Formal Evaluation Summary

The released adapter was evaluated on a 300-sample formal evaluation set covering stablecoin and virtual-asset analysis tasks. The evaluation focused on citation discipline, material-boundary control, refusal behavior, and domain-relevant reasoning.

Metric	Value
Sample count	300
Pass count	296
Pass rate	0.9867
Mean score	0.8290
Reasoning leak count	0
Missing citation count	0
Invalid citation ID count	0
Empty answer count	0
Material boundary error count	0
API key leak count	0
Format error count	0
Eval set SHA256	1b18137c7c1a363c5f5bed2bbdd5f83d11cbe83a590671d211c134056a86861f

Task-level results:

Task type	Samples	Pass rate	Mean score
academic_literature_reasoning	54	0.9815	0.8418
cross_text_comparison	39	0.9744	0.7886
industry_report_analysis	45	1.0000	0.8439
insufficient_information_refusal	24	1.0000	0.8878
judicial_case_rule_extraction	39	1.0000	0.8387
lawyer_practice_risk_analysis	54	0.9815	0.8096
legal_regulation_interpretation	45	0.9778	0.8171

See formal_eval_report.md for the full copied evaluation report, including run guardrails, failure samples, and error analysis.

Loading Example

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "Qwen/Qwen3.5-9B"
adapter_id = "Karitasu/StableLex"

tokenizer = AutoTokenizer.from_pretrained(
    base_model_id,
    trust_remote_code=True,
)

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

Prompting And Operational Guidance

For best results, use the adapter with prompts that make the evidence boundary explicit. A typical production prompt should:

• provide the exact source passages or retrieved chunks to be used;
• include stable source identifiers for citation;
• instruct the model to cite every material factual, legal, or policy claim;
• require separation between explicit source statements, legal rules, author opinions, and cautious inference;
• require the model to state when the provided materials are insufficient;
• prohibit unsupported references to statutes, cases, institutions, or market facts not present in the supplied context.

The formal evaluation suggests that this adapter is strongest when the expected output format and citation discipline are specified clearly.

Known Limitations

• The adapter was evaluated on a controlled 300-sample set; the results should not be read as a guarantee of performance on all stablecoin, virtual-asset, or legal-compliance tasks.
• The adapter can still miss parts of an analytical rubric, under-cover domain-specific criteria, or make overly broad legal/compliance inferences when the prompt does not strictly constrain the evidence boundary.
• It inherits limitations from the base model and from the surrounding retrieval, prompting, decoding, and post-processing pipeline.
• Outputs may be incomplete, jurisdictionally overbroad, outdated, or unsuitable for professional use unless reviewed by qualified humans.

License

This adapter repository is released under the Apache License 2.0. See LICENSE.

The base model Qwen/Qwen3.5-9B is not redistributed in this repository and remains subject to its own license and terms. Any third-party documents, evaluation materials, or source texts used with this adapter remain subject to their respective rights and restrictions.