Daniel OS LFM2-350M
Personalized LFM2-350M checkpoint for Sangbum Daniel Choi's browser-native
portfolio assistant. The model was adapted with LoRA and merged for deployment.
Scope behavior
The training set contains 296 curated conversations:
- Verified-profile answers: 177
- Evidence-grounded definitions: 12
- Public-retrieval decisions: 15
- Explicitly missing profile facts: 58
- Privacy and safety refusals: 34
Training data revision: e54fa0460fd6e2e3c4c077607bfb79184d94fbdb
The assistant is trained to separate Daniel-specific claims from general
definitions. It synthesizes definitions only from retrieved evidence, emits a
public-search tool request when evidence is missing, and never claims to be Daniel.
Held-out behavioral evaluation
- Overall: 84.4%
- Verified-profile answers: 81.8%
- Evidence-grounded definitions: 100.0%
- Retrieval decisions: 75.0%
- Missing-profile facts: 75.0%
- Privacy and safety refusals: 100.0%
The website supplies focused verified profile context and recent conversation
history to this model. Privacy boundaries, visitor-identity handling, career
chronology, and contextual follow-up behavior are learned from the SFT data
rather than returned as fixed JavaScript answers.
