| | --- |
| | title: Math Conjecture Trainer |
| | sdk: gradio |
| | sdk_version: "6.6.0" |
| | python_version: "3.10" |
| | app_file: app.py |
| | pinned: false |
| | emoji: 🧮 |
| |
|
| | --- |
| | |
| | # Math Conjecture Trainer Space |
| |
|
| | An autonomous training operations console for DeepSeek-Math that runs multi-stage curriculum fine-tuning on Space GPU, executes post-training quality evaluation, and publishes only qualified adapters, checkpoints, and run reports to your Hugging Face model repository. |
| |
|
| | This Space is the tactical operations console for `maths-conjuncture-solutions` and is wired to: |
| |
|
| | - dataset: `NorthernTribe-Research/math-conjecture-training-corpus` |
| | - model repo: `NorthernTribe-Research/math-conjecture-model` |
| |
|
| | ## End-to-end flow |
| |
|
| | 1. Download released parquet splits (`train/validation/test`). |
| | 2. Build runtime config from `configs/deepseek_math_sota.yaml`. |
| | 3. Run 4-stage curriculum LoRA fine-tuning with `scripts/train_sota.py`. |
| | 4. Run post-train evaluation (`pass@1`, `pass@k`, exact/boxed, family metrics). |
| | 5. Apply quality gate thresholds before hub push. |
| | 6. Emit `training_summary.json` + `post_eval_report.json` and stream live telemetry in UI. |
| |
|
| | ## Access posture |
| |
|
| | Credentials and publish permissions are handled by deployment runtime settings. |
| |
|
| | ## Runtime controls |
| |
|
| | - `Autonomous Mode`: enabled by default; applies full-stage training/eval/gate/publish profile automatically. |
| | - `Run Evaluation After Training`: toggles post-train eval in runtime config. |
| | - `Enforce Quality Gate`: enables/disables promotion gate checks. |
| | - `Gate Min pass@1`, `Gate Min pass@k`, `Gate Min Rows`: runtime gate thresholds. |
| | - `Live Tactical Telemetry`: real-time stage progression, runtime posture, and training-loss graph (sparkline) with gate/push state. |
| | - `Ops Console (Live Log + Mission JSON)`: unified panel for line-by-line runtime stream, heartbeats, and structured mission summary. |
| | - `Validation Mode (No Training)`: validates pipeline with `--dry-run`. |
| | - `Force Dataset Redownload`: bypasses cached parquet files. |
| | - `Abort Active Run`: cancels active subprocess tree. |
| |
|
| | ## Artifacts |
| |
|
| | - runtime config: `workspace/runtime/deepseek_math_sota.runtime.yaml` |
| | - run output root: `workspace/runs/math-conjecture-sota` |
| | - final adapter: `workspace/runs/math-conjecture-sota/final_adapter` |
| | - training summary: `workspace/runs/math-conjecture-sota/training_summary.json` |
| | - post-eval report: `workspace/runs/math-conjecture-sota/post_eval_report.json` |
| |
|
| | ## Notes |
| |
|
| | - Full training runs on GPU when available and automatically falls back to CPU mode when CUDA is unavailable. |
| | - App handles Gradio copy-button compatibility across versions automatically. |
| |
|
