File size: 2,594 Bytes
47d0cb3
91b3acb
47d0cb3
1165747
 
47d0cb3
 
919f1a4
1a0fccf
47d0cb3
 
91b3acb
 
668cc30
91b3acb
0be512a
91b3acb
0be512a
 
91b3acb
0be512a
91b3acb
0be512a
 
 
 
 
 
91b3acb
1a0fccf
507ee47
1a0fccf
507ee47
0be512a
91b3acb
11f9ebf
0be512a
 
 
a68d3ef
ce9c78b
1a0fccf
0be512a
 
 
 
 
 
 
 
 
 
507ee47
 
 
a68d3ef
0be512a
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
---
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.