galavoxx/llama-3.2-3b-phaedo-ext-comprehensive335

Llama-3.2-3B-Phaedo-Ext-Comprehensive335

Built with Llama

A LoRA fine-tune of Llama 3.2 3B Instruct that reads a conversation and emits structured signals about how the person on one side of it prefers to be communicated with — the extraction stage of Phaedo, a cognitive-fingerprint system that runs entirely on the user's own devices.

What it does

Given a conversation transcript, it emits signals across eight layers (communication style, working style, domain, and so on) in a fixed catalogue. It is not a chat model and it is not useful for general instruction-following — it was trained on one narrow task and is scored on that task alone.

Intended use and limits

• —Intended: local, on-device extraction of communication preferences from the user's own conversations, with the output staying on their device.
• —Not intended: inference about people who have not consented, demographic or protected- attribute inference, or any use the Llama 3.2 Acceptable Use Policy prohibits.
• —Known limit: it emits catalogue signals only. A trait outside the catalogue is not represented, and absence of a signal is not evidence of its opposite.

Training

• —Base: unsloth/Llama-3.2-3B-Instruct
• —Method: SFT (LoRA) via TRL on a synthetic corpus of 335 labelled conversations
• —Evaluation: held-out F1 against a catalogue-anchored gold set, 2026-07-04

Framework versions

• —TRL: 1.5.1
• —Transformers: 5.10.2
• —Pytorch: 2.12.0
• —Datasets: 5.0.0
• —Tokenizers: 0.22.2

Citation

Cite TRL as:

@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}

Quantized build for the browser

galavoxx/llama-3.2-3b-phaedo-ext-comprehensive335-q4f16-MLC is the q4f16 MLC conversion of this model, which is what actually runs in the browser via WebGPU.

License

Llama 3.2 Community License — not an OSI open-source license. The full agreement is in LICENSE.txt in this repository; the required attribution is in NOTICE. Use of this model is subject to the Llama 3.2 Acceptable Use Policy.

Phaedo's own code is Apache-2.0. This model is not, and the distinction matters: a Phaedo deployment that includes this model carries the Llama license terms with it.