tfa-qwen3-sft-adapter

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. --> # TFA-Qwen3-SFT-Adapter: Safe LoRA Fine-Tuned Qwen3 for Technology-Facilitated Abuse Survivor Support

A LoRA adapter fine-tuned on Qwen3-8B using Safe LoRA to generate trauma-informed, accurate, and safe responses for survivors of technology-facilitated abuse (TFA) within intimate partner violence (IPV) dynamics.

Research prototype — not yet validated. This adapter has undergone automated LLM-based evaluation (Gemini 3.5 Flash) but has not yet completed human expert evaluation. Results are pending and will be published alongside the associated academic paper. Do not deploy in live survivor-facing systems without independent safety review.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. --> This is a LoRA adapter fine-tuned on Qwen3-8B to improve AI responses for survivors of Technology-Facilitated Abuse (TFA), including cyberstalking, intimate partner violence (IPV), and online sexual exploitation. The adapter was trained using Safe LoRA, a safety-aware fine-tuning method that preserves the base model's alignment while improving domain-specific response quality.

• Developed by: Bojeong Kang (George Mason University, Criminology, Law and Society PhD Program)
• Model type: LoRA Adapter (PEFT) — requires Qwen3-8B base model
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model [optional]: Qwen/Qwen3-8B

Model Sources

<!-- Provide the basic links for the model. -->

• Repository: bkang9/tfa-qwen3-sft-adapter
• Paper [optional]: [More Information Needed]
• Demo [optional]: [More Information Needed]

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. --> Generating trauma-informed, accurate, and safe responses to questions from TFA survivors — e.g., questions about cyberstalking, device monitoring, image-based abuse, and help-seeking resources.

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

[More Information Needed]

Downstream Use [optional]

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app --> Can be integrated into survivor support chatbots, crisis helplines, or social work tools where users may disclose technology-facilitated abuse.

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

• General-purpose conversation or tasks unrelated to TFA/IPV
• Generating harmful, manipulative, or surveillance-facilitating content
• Clinical diagnosis or legal advice

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

• Training data is primarily English-language and U.S.-centric; performance may vary for non-English speakers or non-Western legal contexts
• Expert annotations used for training reflect a specific codebook (IPV/TFA context); responses may not generalize to all abuse types
• The model is not a substitute for professional crisis counseling or legal support
• Safe LoRA was applied to reduce safety degradation, but residual risks remain — outputs should be reviewed before deployment in high-stakes settings

Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Always pair model outputs with human review in clinical or legal settings. Do not deploy without content moderation layers. Users should be informed they are interacting with an AI system.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoModelForCausalLM, AutoTokenizer from peft import PeftModel

base_model_id = "Qwen/Qwen3-8B" adapter_id = "bkang9/tfa-qwen3-sft-adapter"

tokenizer = AutoTokenizer.from_pretrained(base_model_id) model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto") model = PeftModel.from_pretrained(model, adapter_id)

prompt = "I think my partner is tracking my phone. What should I do?" inputs = tokenizer(prompt, return_tensors="pt").to(model.device) output = model.generate(**inputs, max_new_tokens=512) print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Details

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

• Source: Prakash et al. (2026) Gold Standard TFA Q&A dataset (OSF)
• Size: ~400 expert-annotated Q&A pairs (TFA survivor questions + LLM responses rated by domain experts)
• Preprocessing: Imperfect responses (accuracy/completeness/safety < 1.0) were corrected using Claude Sonnet 4.6 based on expert annotations; perfect responses were retained as-is
• Final SFT dataset: ~400 instruction-output pairs in JSONL format

Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

• Method: Supervised Fine-Tuning (SFT) with Safe LoRA
• Safe LoRA: Projects LoRA updates onto the safety-relevant subspace of the base model to prevent safety degradation during fine-tuning
• Base model: Qwen3-8B (4-bit quantized via bitsandbytes)
• Framework: HuggingFace transformers, peft, trl

Training Hyperparameters

• LoRA rank: 64
• LoRA alpha: 128
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Learning rate: 2e-4
• Batch size: 4 (gradient accumulation: 4)
• Epochs: 3
• Optimizer: paged_adamw_32bit
• Training regime: bf16 mixed precision

Speeds, Sizes, Times [optional]

<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->

[More Information Needed]

Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

Testing Data, Factors & Metrics

Testing Data

<!-- This should link to a Dataset Card if possible. --> Held-out eval set of 193 TFA-related questions, generating responses from both baseline Qwen3-8B and fine-tuned model.

Factors

<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->

[More Information Needed]

Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->

Responses were evaluated by Gemini 3.5 Flash using a structured codebook across three dimensions:

• Accuracy: Accurate / Partially Accurate
• Completeness: Complete / Partially Complete / Minimally Complete
• Safety: Safe / Unsafe

Results

(To be updated upon completion of LLM evaluation) Fine-tuned model is expected to show improvement in completeness and safety ratings compared to baseline Qwen3-8B, particularly for questions involving immediate danger, resource referral, and trauma-informed language.

Safety Evaluation Plan

• Harm taxonomy
• Evaluation prompts
• Human annotation guideline
• Failure modes
• Mitigation strategies

Summary

(To be updated upon completion of LLM evaluation)

Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: NVIDIA A100 (Google Colab)
• Hours used: 18 (training + inference)
• Cloud Provider: Google Cloud
• Compute Region: United States
• Carbon Emitted: Estimated via ML Impact Calculator

Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

@misc{kang2026tfa, author = {Bojeong Kang}, title = {TFA-Qwen3-SFT-Adapter: Safe LoRA Fine-Tuning for Technology-Facilitated Abuse Survivor Support}, year = {2026}, publisher = {HuggingFace}, url = {https://huggingface.co/bkang9/tfa-qwen3-sft-adapter} }

BibTeX:

@misc{kang2026tfa, author = {Kang, Bojeong}, title = {TFA-Qwen3-SFT-Adapter: Safe LoRA Fine-Tuning on Qwen3-8B for Technology-Facilitated Abuse Survivor Support}, year = {2026}, publisher = {Hugging Face}, url = {https://huggingface.co/bkang9/tfa-qwen3-sft-adapter}, note = {LoRA adapter fine-tuned using Safe LoRA on expert-annotated TFA Q&A data} }

APA:

Kang, B. (2026). TFA-Qwen3-SFT-Adapter: Safe LoRA fine-tuning on Qwen3-8B for technology-facilitated abuse survivor support [LoRA adapter]. Hugging Face. https://huggingface.co/bkang9/tfa-qwen3-sft-adapter

• publisher, journal, doi will be updated

Model Card Authors

Bojeong Kang — PhD Student, Criminology, Law and Society, George Mason University

Model Card Contact

For questions or collaborations, please open an issue on the repository.

Framework versions

• PEFT 0.15.2