Measured product lead custody receipt.
github.com/Zer0pa/Polymath-AIRESEARCH-READY
LIVE
PHONE COMPILE OPEN
Polymath LLM Training On Mobile Phone
On-device language-model training research lane · Polymath-AI · PyPI 0.1.0 · Snapdragon 8 Elite target
Polymath-AI is a training harness aimed at the Snapdragon 8 Elite (SM8750) phone chip. It trains only the first and last layers of a language model while the middle stays sealed and SHA-checked. The host smoke runs cleanly on Qwen 2.5 1.5B with the frozen middle showing zero weight changes.
Phone compilation, licensed multilingual corpora, sustained device telemetry, and a public checkpoint are all open. This is a route, not a product.
01 · THE GAPPHONE RUN MISSING
“Training a language model on a phone still has no measured path from corpus to battery.”
02 · MARKETSUSER FIT
Best fit is the research-infrastructure and mobile-runtime audience deciding what to staff; no model-revenue claim is made.
03 · VALUE
OPENNOW
Public repo and PyPI exist; the value is the training harness and its constraints, not a phone-trained model.
04 · INSIGHT
A training harness, not a finished model.
05.0 · CURRENT TECHHOST, CPU, NATIVE
Mobile language-model work usually means inference on the chip, with training kept in the cloud. The conventional route ships a trained model down to the device and never lets it learn there.
05.1 · OUR TECHSELECTIVE LAYER TRAINING
Polymath trains only the boundary layers of a language model — layer 0, the final layer, and the language-model head — while every middle layer stays sealed and SHA-checked at frozen_changes = 0. Host smoke passes on Qwen 2.5 1.5B, with loss falling from 14.515 to 4.449 in five steps and the middle bit-identical across the run.
05.2 · BENCHMARKSHOST HARNESS
Host testsPASSreported host
Smoke baseQwen 2.51.5B params
Checks19listed
SoCSM8750SD 8 Elite resolved
Device status: five SM8750 phone-compile rows currently measured unsupported; host harness passes.
06 · MEASUREMENTHOST ELO SMOKE
Host smoke passes, phone compile remains unsupported.
06.1 · COMPARATIVE PERFORMANCE · HOST VS DEVICE STATUS
Host smoke · Qwen 2.5 1.5B, 5 training steps, loss 14.515 to 4.449, frozen middle unchanged. Phone compile, licensed corpus ingestion, and sustained device telemetry are not yet measured.
07 · KEY METRICSPOLYMATH-AI HOST HARNESS · PYPI 0.1.0 STALE
07.1 · HOST TEST SURFACE
PASS
Host harness pass · reported on developer machine
07.2 · SMOKE BASE
Qwen 2.5·1.5B
Smoke base model · frozen_changes = 0
07.3 · CHECK ROWS
19
Listed status rows · documentation coverage
07.4 · TARGET SOC
SD 8 Elite·open
SM8750 resolved · phone compile blocked
07.5 · ON-DEVICE THROUGHPUT
null
Metric absent · device path unsupported
08 · DETERMINISMFROZEN MIDDLE · SHA-CHECKED
Frozen middle stays bit-stable while boundary layers train.
08.1 · WHAT DETERMINISTIC MEANSFROZEN_CHANGES = 0
Only the named boundary layers receive gradient updates — layer 0, the final layer, and the language-model head. The middle layers' weights are SHA-checked before and after every training pass; if any frozen weight moves, the run halts immediately and reports the offending tensor.
The unit of bit-exactness is per-pass, host-side. Five steps on Qwen 2.5 1.5B leave the frozen middle unchanged across the entire run. No on-device determinism claim is made yet; the Qualcomm Neural Network and LiteRT paths are not exercised.
08.2 · THE FIDELITY GAP
Honest Blocker ·
QNN/LiteRT compile on the Snapdragon 8 Elite is measured unsupported, so the scheduler cannot reach the device yet. On-device execution, sustained telemetry, licensed-corpus ingestion, and the next PyPI release all remain open. Tokenization currently bloats Zulu 2.68× and Greek 4.38× past target. No phone-trained model or public checkpoint exists.
09
FIVE PATHS FROM ONE phone-side training loop.
09.1 · THIS REPO'S AMBITION
The hinge is selective continual pretraining under real mobile constraints. Polymath-AI does not promise a finished model. It builds the scheduler, corpus discipline, and frozen-middle guarantee needed to answer one question honestly — whether training a useful language model on a phone, under battery and thermal limits, is worth doing at all.
09.2 · WHAT WORKS NOW
Working now: host training harness on Qwen 2.5 1.5B, frozen-middle SHA-check, scheduler framing, and a resolved chip target.
09.3 · WHAT'S STILL OPEN
Still open: phone compile path, sustained device telemetry, licensed multilingual corpora, and a published checkpoint with release evidence.
09.4 · ADAPTATION · NEAR-TERM (12–24 MO)
The fine-tune leaves the data centre
A mobile-runtime engineer who can land a boundary-layer training pass on a flagship chip stops needing a remote fine-tune to personalise a model. Adaptation becomes a battery decision on the device, not a procurement decision with a cloud vendor.
09.5 · CORPUS CUSTODY · NEAR-TERM (12–24 MO)
Multilingual data stops travelling
When the training step runs on the handset, the multilingual text a model learns from no longer has to leave the phone. A speaker of an underrepresented language can contribute to their own model without their words crossing a corporate boundary.
09.6 · PERSONAL MODELS · MID-TERM (24–48 MO)
One model, one person, one phone
If selective training holds at scale, a model can drift toward the person carrying it rather than the average of millions of strangers. The phone becomes a place where a small, personal model improves over months instead of being replaced quarterly.
09.7 · RECEIPTS · MID-TERM (24–48 MO)
Mobile training answers to evidence
A regulator or platform reviewer who asks how an on-device model changed can be answered with a record — layer touched, update size, battery cost, quality movement — rather than a marketing claim. Phone training becomes something assessable, not just demonstrated.
09.8 · LOCAL AGENCY · PARADIGM (48 MO+)
The phone becomes a knowledge instrument
Once training, telemetry, and corpus custody all fit inside the device, the phone stops being the last mile of someone else's model. It becomes a bounded place where a person's language, history, and tasks shape what their model knows.