LoRA fine-tunes that impersonate dhilipsiva — they ARE his website: served into the visitor's browser and run entirely on-device via candle compiled to WebAssembly.
--- license: apache-2.0 language: [en] base_model: - HuggingFaceTB/SmolLM2-135M-Instruct - Qwen/Qwen2.5-0.5B-Instruct tags: [gguf, persona, candle, webassembly, chatml] --- # dhilipsiva-twin — on-device persona models LoRA fine-tunes that impersonate [dhilipsiva](https://dhilipsiva.dev) — they ARE his website: served into the visitor's browser and run entirely on-device via [candle](https://github.com/huggingface/candle) compiled to WebAssembly. | file | base | size | extra trick | |---|---|---|---| | `dhilipsiva-twin-q8_0.gguf` | SmolLM2-135M-Instruct | 145MB | persona | | `dhilipsiva-twin-qwen-q8_0.gguf` | Qwen2.5-0.5B-Instruct | 531MB | persona + emits `TOOL {"app":...}` lines that open the site's MCP apps | Tokenizers included as `tokenizer-smol.json` / `tokenizer-qwen.json`. ChatML prompting. **The system prompt must match the training prompt verbatim** — see `finetune/generate_dataset.py` in the [site repo](https://github.com/dhilipsiva/dhilipsiva.com) (`SYSTEM` for smol, `SYSTEM_TOOLS` for qwen). Low-temperature decoding recommended (temp ~0.3): they answer *as* dhilipsiva on questions about him, and answer general questions plainly in his voice — fit with a contrast corpus so they no longer recite his bio for every prompt. ⊥ **These models will lie, confidently.** Fluent ≠ true — that gap is the point: it's why dhilipsiva builds [nibli](https://github.com/dhilipsiva/nibli), a hallucination firewall that derives answers with proof traces instead of predicting plausible text. Trained facts are accurate as of 2026-06; everything else is improv.
LoRA fine-tunes that impersonate dhilipsiva — they ARE his website: served into the visitor's browser and run entirely on-device via candle compiled to WebAssembly.
