Fix Agron training pipeline and improve response quality
Njap-png
d9ba17d
12 days ago
Add Agron neural-network zoo, optimisers, algorithms, schedulers and unified trainer
Njap-png
8922eae
- Implement 10 architectures (native transformer/MoE/hyper, hybrid/mamba SSM, GRU/LSTM, causal conv, Mixer) with HF-compatible save/load/generate
- Add 16 optimisers (Lion/Muon/Shampoo/Adan/LAMB/Sophia/ScheduleFree/Kahan/...)
- Add 16 composable training algorithms + 8 schedules + 10 training recipes
- New agron_nets.training engine with gradient clipping/accumulation/SAM/EMA/Z-loss/composability
- Wire --arch/--optimizer/--scheduler/--algorithm/--recipe into train_from_scratch.py and train_agron.py
- Upgrade app.py with zoo UI, architecture switching and engine-based training
- Add full test-suite (110 tests) and smoke test across all architectures
Clear test interaction log
Njap-png
1c16fcb
Polish Gradio HUD, drop mock dashboard, prep Space deploy
Njap-png
51e553e
- Remove agron_dashboard.html: it was a fully simulated mock (canned
replies, random metrics) and its 92vh iframe was pushed out of view
by the old `html,body{height:100vh;overflow:hidden}` rule, so it was
removed rather than embedded.
- Re-enable normal page scrolling and drop the duplicated pseudo-header
branding; add responsive rules for narrow viewports.
- Restructure the UI into COMMAND CONSOLE + SYSTEM/TOOLS/LEARNING tabs so
the control column can no longer overflow off-screen.
- Split the status box into a live TELEMETRY strip and a persistent
ACTION LOG, so button feedback is no longer overwritten every 5s.
- Replace hardcoded placeholder metrics with real model/device/GPU state.
- Fix None crashes: web_topic.value, tool description and chat message
could all be None and raise AttributeError.
- Skip the per-message web search on the base-model path (it was fetched
and discarded) and return a multi-source grounded answer instead.
- Target Hugging Face, not GitHub…
19 days ago
Add factory reset (fresh model + memory)
Njap-png
9f66945
Clean learned topics and persist deduped knowledge
