exnivo / Talky-GPT2-SFT

Talky GPT-2 124M SFT

Talky GPT-2 124M SFT is a small GPT-2 based conversational model fine-tuned on the Talky conversational dataset. It is designed for short, casual, friendly chat replies with a relaxed “vibe bot” style.

This model is mainly an experiment in supervised fine-tuning for personality and conversational style. It is not intended to be a factual assistant or reasoning model.

Model Details

• Model name: Talky GPT-2 124M SFT
• Base model: GPT-2 style / GPT-2 small architecture
• Approx. size: ~124M parameters
• Training method: Supervised Fine-Tuning (SFT)
• Dataset: Talky conversational JSONL dataset
• Training examples: 58,748
• Format: User: / Assistant: prompt-completion format
• Language: English
• License: MIT

Intended Use

This model is intended for:

• Short casual conversations
• Friendly chatbot experiments
• Joke-style responses
• Game NPC dialogue
• Roblox/chatbot prototypes
• Style-transfer experiments for small language models

It works best when used as a lightweight conversational model, not as a factual or coding assistant.

Limitations

This model is small and should not be expected to perform well on reasoning, math, factual questions, coding, or real-time information.

Known limitations:

• May answer math incorrectly
• May hallucinate facts
• May not know the current time or current events
• May dodge technical questions
• May repeat patterns from the training data
• May produce casual replies even when a factual answer is needed

Safety and Responsible Use

This model is experimental. Outputs should be monitored if used in public chat applications, games, or apps for younger users.

Do not use this model for:

• Medical advice
• Legal advice
• Financial advice
• Safety-critical decisions
• Factual knowledge without verification
• Math or coding tasks where correctness matters

Notes

This model is best viewed as a small personality/style experiment. It can generate casual, funny, friendly replies, but it does not have strong reasoning ability.

Citation / Credits

Fine-tuned by Guus / AI Nexus Studios as a small conversational SFT experiment.