Datasetpublic

Tribewarez/synthetic-pot-o-challenges-v1

synthetic-pot-o-challenges-v1 Tiny synthetic starter dataset for training PoT-O (Proof of Tensor Optimizations) pathfinder models. Format (JSONL) {"challenge": "tensor:shape=[32,64];dtype=float16;target_mml=0.42;ops:matmul,lowrank,gelu,quant4,prune0.3,transpose", "optimal_path": "path: matmul[lowrank:16] -> relu -> quant:4bit -> prune:0.35 -> score:0.418"} challenge: Text encoding of tensor properties & allowed operations optimal_path: Heuristic "good"… See the full description on the dataset page: https://huggingface.co/datasets/Tribewarez/synthetic-pot-o-challenges-v1.

sourceHugging Facemitupdated 7mo agoView on Hugging Face

0likes17downloads

Dataset Card

synthetic-pot-o-challenges-v1

Tiny synthetic starter dataset for training PoT-O (Proof of Tensor Optimizations) pathfinder models.

Format (JSONL)

json

{"challenge": "tensor:shape=[32,64];dtype=float16;target_mml=0.42;ops:matmul,lowrank,gelu,quant4,prune0.3,transpose", "optimal_path": "path: matmul[lowrank:16] -> relu -> quant:4bit -> prune:0.35 -> score:0.418"}

—challenge: Text encoding of tensor properties & allowed operations
—optimal_path: Heuristic "good" optimization sequence + predicted score (MML-inspired efficiency)

Dataset Details

—Size: 100 examples (v1 – expand in future versions)
—Train/Val Split: 90/10
—Generation: Rule-based synthetic (random shapes/dtypes/targets + simple heuristics for paths). Not from real traces yet.
—Intended use: Fine-tuning tiny models (e.g. Tribewarez/pot-o-pathfinder-tiny-v1) to predict better tensor transformation paths for low-power PoT-O miners.

Next Iterations

—Add real tensor traces from ai3-lib
—More diverse challenges
—Verified optimal paths via solvers
—500+ examples with varied op combinations
—Create v2 with real matrix compression benchmarks

MIT licensed • Tribewarez guild • Live beta • 2026

Link to Model

In your model README.yaml add:

yaml

datasets:
  - Tribewarez/synthetic-pot-o-challenges-v1