Configuration Parsing Warning:In adapter_config.json: "peft.task_type" must be a string
PhonoCommand LLM 0.8B
User-selected current Web default: Qwen3.5-0.8B + single r16 LoRA + a learned candidate head. This repository contains adapter weights and head.safetensors; the pinned official base is downloaded separately. It is not a standalone chat model and must not be used as generic autoregressive/Ollama generation.
One backbone forward classifies command Y/N and ranks supplied ordered I/T/P candidates. Internal unresolved None differs from the business string 'none'. No extra system prompt, string fixes or confidence guard. Maximum 512 tokens. Continuation step0 inherits trained step4400; it is not untrained. This release is deliberately the current Web default, not the separate step700 experiment.
ALGORITHM.txt explains the token layout and decision head. WINDOWS_HANDOFF.txt explains native Windows integration. reference_code.zip contains the small parser and example. Use the exact raw base commit in upstream_lock.json, merge one adapter, then load the head. Pass the downloaded local base directory explicitly to the parser. Use warm_lengths=() for eager SDPA; CUDA graphs are an optional optimization. Keep candidate order and tokenization unchanged. Supply the target app's actual command taxonomy rather than copying the old Web project's candidate lists.
Selected checkpoint validation: unseen utterances 1651/1653; natural Cantonese 318/320; mixed 1193/1200; reordered mixed 1195/1200. These are checkpoint selection metrics, not an independent Windows/robot test. No perfect-score claim. Weights are unchanged in export; only private path metadata is normalized. Apache-2.0. No dataset rows or human audio are included.
- Downloads last month
- -