Hinglish Smooth Talker — Experimental
I tried teaching a 3B model to flirt in Hinglish. It loses the plot after a few turns.
The idea: a smooth-talker persona for Bumble, Tinder, and Hinge — natural Hindi-English conversation, good banter, and bios with some personality. The current model is bad at sustaining a conversation. It can drift into nonsense after a few turns.
I'm publishing the attempt, including the weights and training details, while I figure out how to make it better. Unfinished experiment. No claims of real-world success.
Goals
- Fluent, natural Hindi-English conversation.
- Contextual openers, playful replies, and flirting that fits the conversation.
- Fun dating bios and profile-prompt answers.
These are goals, not capabilities this checkpoint reliably delivers.
Can it survive five turns?
Try a short Hinglish conversation and see where it breaks. If you find a useful reply or a spectacular failure, share the prompt, output, and settings in Community. Use fictional or anonymized chats.
I'm especially interested in fixing repetition, awkward Hinglish, and losing context. Comparisons with the base model are welcome. This is an invitation to test, not a benchmark result.
Current checkpoint
Qwen2.5-3B-Instruct · MLX 4-bit · LoRA · Mac mini M4
This release contains v3, checkpoint 4,000 from a 6,000-iteration run using 70,730 filtered training conversations. Included: the saved fused model (~1.74 GB), selected adapter, training configuration, and an inference script. No formal quality evaluation has been completed.
The training mixture includes Hinglish, English, flirting, banter, and instructions. Sources: FlirtFlip, Flirty-Dialouge, the-rizz, hinglish-conv-dataset, Sujalvc, Roshan32, and chad-sft. Training details and source links.
Local inference
Requires an Apple silicon Mac.
pip install 'mlx-lm==0.29.1' huggingface_hub
hf download inglebhavin98/hinglish-smooth-talker-qwen2.5-3b-experimental --local-dir ./smooth-talker
cd smooth-talker
python examples/chat.py
Planned work
Later, I want to build a harness with conversation memory and a user-defined goal, such as suggesting a date or asking to exchange numbers when interest is mutual. The harness is not implemented in this release. It would keep the user in control of messages and respect the other person's boundaries.
Built with Qwen. Qwen Research License · Attribution. Non-commercial research/evaluation; commercial use requires a separate upstream license.