A Paranormal Investigator LLM — fine-tuned from Qwen2.5-7B-Instruct on a curated corpus of Indian paranormal encounter transcripts, investigator blogs, and first-person supernatural experience accounts.

Designed to respond like an experienced paranormal investigator: open-minded yet analytical, fluent in Hinglish, and respectful of both scientific inquiry and lived supernatural experiences.

Model Capabilities

🎙️ Hinglish-first: Responds naturally in Hindi-English code-mixing, matching how Indian paranormal communities communicate
🔍 Investigator persona: Balances skepticism with belief — investigates claims seriously without dismissing or fear-mongering
🏠 Indian context: Understands local ghost lore (bhoot-pret, tantra-mantra, aatma), regional beliefs, and culturally relevant advice
📖 Storytelling: Can narrate Indian folklore and paranormal encounter stories in an engaging first-person investigator style

Example Interaction

User: Maine ek raat ko ghar me ajeeb awaaz suni. Kya ho sakta hai?
Ghosty: Aise ajeeb awaaze ki koi common reasons hote hain:
1. Electrical interference (like faulty appliances)
2. Humming from AC/ventilator
3. Koi animal (katra, bat)
4. Wind noise through windows
...

User: Kya bhoot sach me hote hain?
Ghosty: Bhoot ka concept India ke folklore aur spiritual traditions mein bahut purana hai,
lekin modern scientific perspective se aise entities ko "unexplained phenomena" ya
"paranormal activity" ke roop mein label kiya jata hai...

Training Data

Trained on a diverse collection of:

Paranormal investigator blogs & transcripts (Creative Commons licensed, used with permission)
Indian ghost encounter narratives from verified public sources
First-person supernatural experience accounts from Indian online communities
Hinglish conversational data spanning urban and rural paranormal reporting styles

Dataset: ~3,000 SFT question-answer pairs in Qwen chat format.

All data sourced from publicly available, permissively licensed Indian paranormal content.

Training Details

Parameter	Value
Base Model	Qwen2.5-7B-Instruct (abliterated variant)
Fine-tuning	Full parameter FT (not LoRA)
Hardware	NVIDIA H100 80GB
Precision	BF16
Attention	SDPA (PyTorch built-in)
Epochs	5
Batch Size	8 (effective 16 with grad accum)
Learning Rate	2e-5 (cosine, warmup 185 steps)
Max Length	2048 tokens
Gradient Checkpointing	Yes

Intended Use

Ghosty-7B is designed for:

Paranormal investigation RP/chatbots
Indian folklore & ghost story generation
First-responder paranormal advice systems
Educational demonstrations of Indian supernatural beliefs

Limitations

May occasionally hallucinate specific case details (standard for LLMs)
Best at Hinglish — performance in pure English or pure Hindi may vary
Not a replacement for professional mental health or law enforcement advice
Trained on publicly shared experiences — individual accounts may not be verified

License

Apache 2.0

saidutta69/Ghosty-7B

Ghosty-7B 👻