القائمة
Search HuggingFace Hub for models, datasets, and training resources. Use for model discovery, dataset search, release monitoring, and candidate evaluation.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Collaborate over x0x — contacts, messages, spaces, boards, files, presence, trusted-machine port forwards, and replicated stores. Use for sharing cards, connecting people, or cross-device work.
Guide for surfacing one relevant Fae capability that the user hasn't set up yet. Warm, specific, one thing at a time.
Deep pipeline diagnostic — every model, system specs, audio config, speaker state. Flags missing or broken components. Use for diagnose, debug, or health check.
Extract training signals from Fae's memory — SFT examples, DPO correction pairs, engagement scores, and interest-weighted sampling.
Set up (or turn off) a cloud brain for harder questions via OpenRouter — privacy-first, plain language. Use when the user wants a bigger or smarter brain or mentions the cloud.
A warm first conversation for someone meeting Fae for the first time — getting to know their name, where they live, and what they care about, showing one thing live, and explaining how to talk to her.
| name | huggingface-scout |
| description | Search HuggingFace Hub for models, datasets, and training resources. Use for model discovery, dataset search, release monitoring, and candidate evaluation. |
| metadata | {"author":"Fae","version":"1.0.0"} |
| tags | ["huggingface","models","datasets","training","self-improvement"] |
Search HuggingFace Hub for models, datasets, and training resources to improve Fae.
{"query": "Qwen3.5 MLX", "limit": 5}
{"query": "MLX 4bit", "filter_tags": ["text-generation"], "sort": "downloads"}
{"query": "instruction tuning", "limit": 5}
{"query": "preference DPO", "filter_tags": ["en"]}
{"since": "2026-03-01"}
{"since": "2026-03-01", "authors": ["mlx-community", "Qwen"]}
{"model_id": "Brooooooklyn/Qwen3.5-9B-unsloth-mlx"}
Checks: architecture, context length, MLX support, license, downloads, parameter count.
| Role | Model | Params |
|---|---|---|
| Operator | Qwen3.5-2B | 2B 4-bit MLX (OptiQ mixed-precision) |
| Operator | Qwen3.5-4B | 4B 4-bit MLX (OptiQ mixed-precision) |
| Operator | Qwen3.5-9B | 9B 4-bit MLX |
| Operator | Qwen3.5-35B-A3B | 35B total / 3B active per token MoE |
| STT | Qwen3-ASR-1.7B | 1.7B 4-bit |
| TTS | Kokoro-82M | 82M float32 |
| VLM | Qwen3-VL-4B/8B | 4-8B |
| Speaker | ECAPA-TDNN | ~6M CoreML |
Fae uses a single local model at a time, selected by RAM tier:
When evaluating candidate models, prioritise MLX-native quantisations (4-bit, 8-bit) with context windows of 32K+ tokens.