Math Conjecture Trainer Space

Autonomous Space Trainer

This app runs the full training pipeline for the maths-conjuncture-solutions stack:

Pull parquet splits from NorthernTribe-Research/math-conjecture-training-corpus.
Build runtime config from configs/deepseek_math_sota.yaml.
Execute staged LoRA curriculum via scripts/train_sota.py.
Run post-training evaluation and read quality gate outcomes.
Push artifacts only when policy/gates allow.
Stream live operations telemetry in a monochrome command-center UI.

Autonomous and continuous modes are enabled by default.

Runtime Controls

Dataset Source

Model Destination

Base Model ID

Autonomous Mode

Continuous Auto-Restart (Enforced)

Stage Start

1 4

Stage Count

1 4

Run Evaluation After Training

Evaluation K

1 8

Evaluation Max Samples

50 1000

Enforce Quality Gate

Gate Min pass@1

0 0.5

Gate Min pass@k

0 1

Gate Min Rows

10 2000

Push Adapter to Hub

Force Dataset Redownload

Validation Mode (No Training)

Live Outputs

Run Status

Run Log (Live Log + Summary JSON)