Instructions to use allnamestaken177/grimoire-volatility-laya with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Laya
How to use allnamestaken177/grimoire-volatility-laya with Laya:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Laya volatility checkpoint for Grimoire
A fine-tune of Convai's Laya (ModernBERT-large, 421M parameters) that answers two questions about one fact from an AI agent's memory:
changes(yes/no): does it describe current state that can go out of date, rather than a decision, rule, historical event or finding?speed(hours / days / weeks / months / years): if it changes, how fast?
Grimoire asks this once when a fact is written and uses the answers to decide when recall should tell an agent to re-check the fact. See docs/FRESHNESS.md.
Use
pip install laya
LAYA_EXTRA_MODELS='{"volatility": "<path to this repo>"}' LAYA_DEFAULT_MODEL=volatility laya-serve
export GRIMOIRE_DECISION_URL=http://127.0.0.1:8000
export GRIMOIRE_DECISION_MODEL=volatility
export GRIMOIRE_DECISION_CALIBRATION=2.147,0.359
Ask the questions exactly as in questions.json.
Training data
train.jsonl: 899 synthetic facts written by a small LLM to cover live
state, work status, locations, configuration, decisions, rules and
findings. Both questions were labelled as soft probabilities by TypeSafe's
Jev. No real user memory is in the training data.
laya-train --loss soft-ce --shuffle-options --epochs 6 --max-len 256, CPU.
Evaluation
On 90 hand-labelled facts from a real agent memory store (25 describe changing state). None of them appears in the training data.
| Predictor | AUC | Speed right (of 25) |
|---|---|---|
| Stock Laya, zero-shot | 0.50β0.55 | β |
| Jev | 0.90 | 14 |
| This checkpoint | 0.92 | 15 |
On 90 facts that is a tie with Jev, not a win: the 95% bootstrap interval for the AUC gap is [β0.06, +0.11]. The Platt calibration above was fitted on the same 90 facts. Refit it on your own labelled facts if you can.
Limits
One person's agent-memory style, English only, with code- and ops-shaped facts. Probabilities run high before calibration.
- Downloads last month
- 4
Model tree for allnamestaken177/grimoire-volatility-laya
Base model
convaiinnovations/laya