34 lines
1.2 KiB
Markdown
34 lines
1.2 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
base_model: HuggingFaceTB/SmolLM2-135M-Instruct
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
tags:
|
||
|
|
- safetensors
|
||
|
|
- study-assistant
|
||
|
|
- fine-tuned
|
||
|
|
---
|
||
|
|
|
||
|
|
# Wyzer 2.0 (SmolLM2-135M fine-tune)
|
||
|
|
|
||
|
|
The conversational companion inside the **BackOnTrack** study app.
|
||
|
|
LoRA fine-tune of SmolLM2-135M-Instruct, merged to native Transformers
|
||
|
|
format (safetensors) — this is the deployment artifact served by the
|
||
|
|
Wyzer FastAPI Space. An optional `wyzer-2.0-q4_k_m.gguf` is included for
|
||
|
|
local llama.cpp inference. The 135M size was chosen deliberately: it
|
||
|
|
trains on a laptop CPU and responds fast on Hugging Face's free 2-vCPU
|
||
|
|
tier.
|
||
|
|
|
||
|
|
Trained behaviours:
|
||
|
|
|
||
|
|
- Grounded answers: user metrics arrive in a `[CONTEXT]...[/CONTEXT]` block;
|
||
|
|
the model quotes numbers only from there.
|
||
|
|
- Refusal to fabricate: missing data → says so instead of guessing.
|
||
|
|
- Action honesty: recommends, never claims to execute; the app's
|
||
|
|
deterministic engine performs approved actions.
|
||
|
|
- Mentor persona across study, tech, career, fitness, wellbeing (English +
|
||
|
|
Hinglish).
|
||
|
|
|
||
|
|
Trained offline with a synthetic instruction dataset (~900 examples,
|
||
|
|
2 epochs, LoRA r=16). Not a general-purpose assistant; expect app-domain
|
||
|
|
behaviour.
|