Independent use will determine whether the released Deno/WebGPU trainer can practically train tiny language models directly in GGUF while producing checkpoints reliably compatible with llama.cpp.
state: expiredheat: lowuncertainty: highconvergesscott: mediumlocal-inference open-model-training ggufFelladrin
What is this?
Felladrin announced an open-source TypeScript trainer intended to pretrain tiny language models from scratch and write them directly to GGUF without Python, using Deno/WebGPU according to the case. The supplied web answer claims its checkpoints work with llama.cpp and that training can run across varied hardware, including CPUs. However, the search snippets only establish GGUF as a format used by llama.cpp for local inference; they do not independently verify the trainer’s reliability, hardware coverage, or checkpoint compatibility, so those claims still await external reproduction.
Why it matters to Scott
A reliable direct-to-GGUF trainer would extend Scott’s existing local-model infrastructure and synthetic training-data work from inference/data preparation into portable model creation, while supporting his Model Perishability preference for swappable, non-vendor-bound artifacts. The opportunity is actionable but still conditional: compatibility and training reliability have not been independently reproduced, and the stated scope is only tiny models.
ip:concept.model-perishabilitydev:technology.ollamadev:concept.synthetic-finetuning-datasetdev:concept.hardware-aware-local-inferenceradar:concept.ggufradar:concept.webgpuradar:concept.llama-cppradar:gguf-lora-16gb-moe-training
queries asked of Scott's wikis
- direct-to-GGUF training pipelines
- WebGPU or browser-native model training
- TypeScript and Deno for AI tooling
- tiny language models trained locally
- llama.cpp checkpoint interoperability
- open-model training without Python
Measured heat
no measured readings yet — the hourly heat pass fills this in
How the heat travelled
no chain yet — the hourly chain pass fills this in
Evidence (2) — ⭐ canonical anchor
Interpretation history
2026-08-24T05:22:25Z
Repeated checks have found no independent training run, llama.cpp compatibility test, or adoption; the modest engagement increase adds no substantive validation, so active monitoring is no longer justified.
2026-08-22T04:28:27Z
The staleness check adds no independent reproduction, llama.cpp load test, or adoption signal; the case remains a dormant first-party experiment rather than an emerging direct-to-GGUF workflow.
2026-08-20T04:22:15Z
No independent training run, checkpoint-load test, or downstream adoption has appeared, so the release remains an unvalidated first-party experiment rather than an emerging workflow.
2026-08-18T03:29:15Z
The reevaluation adds no independent reproduction or compatibility testing, so the trainer remains a promising first-party release whose practical direct-to-GGUF workflow is unvalidated.
2026-08-18T03:28:34Z
grounded: converges/medium — A reliable direct-to-GGUF trainer would extend Scott’s existing local-model infrastructure and synthetic training-data work from inference/data preparation into
2026-08-18T03:25:49Z
origin walked (codex/luna, conf 0.98): anchor reddit.post.1vrd9lz -> echo.github.c36cc26894 by Victor Nogueira (Felladrin)
2026-08-18T03:23:20Z
case created — The author reports a concrete open-source trainer and 95M checkpoint enabling a novel direct-to-GGUF local training workflow.
Decision trace
- 08-24 15:22expireRepeated checks have found no independent training run, llama.cpp compatibility test, or adoption; the modest engagement increase adds no substantive validation, so active monitoring is no longer just
- 08-24 15:22alert_silentOnly staleness and minor engagement changed; there is no consequential new evidence to surface. A future independent reproduction can reopen the episode.
- 08-24 15:22alert_routeOnly staleness and minor engagement changed; there is no consequential new evidence to surface. A future independent reproduction can reopen the episode.
- 08-22 14:28repriceThe staleness check adds no independent reproduction, llama.cpp load test, or adoption signal; the case remains a dormant first-party experiment rather than an emerging direct-to-GGUF workflow.
- 08-22 14:28alert_silentNo consequential delta has occurred since the prior review. It can wait unless an independent user reports a successful training run and verifies that resulting checkpoints load correctly in llama.cpp
- 08-22 14:28alert_routeNo consequential delta has occurred since the prior review. It can wait unless an independent user reports a successful training run and verifies that resulting checkpoints load correctly in llama.cpp
- 08-20 14:22repriceNo independent training run, checkpoint-load test, or downstream adoption has appeared, so the release remains an unvalidated first-party experiment rather than an emerging workflow.
- 08-20 14:22alert_silentThe staleness trigger adds no consequential evidence; this can wait until an independent user demonstrates successful training and llama.cpp compatibility.
- 08-20 14:22alert_routeThe staleness trigger adds no consequential evidence; this can wait until an independent user demonstrates successful training and llama.cpp compatibility.
- 08-18 16:21sensor_dirtyengagement_update
- 08-18 13:29repriceThe reevaluation adds no independent reproduction or compatibility testing, so the trainer remains a promising first-party release whose practical direct-to-GGUF workflow is unvalidated.
- 08-18 13:29alert_silentThere is no new consequential delta beyond the already assessed release; wait for an independent successful training run and llama.cpp checkpoint load.
- 08-18 13:29alert_routeThere is no new consequential delta beyond the already assessed release; wait for an independent successful training run and llama.cpp checkpoint load.
- 08-18 13:28alert_silentThe public release is established and technically relevant, but its practical value depends on reproducing WebGPU training reliability and llama.cpp-compatible checkpoints. With only tiny-model scope
- 08-18 13:28surface_candidateThe public release is established and technically relevant, but its practical value depends on reproducing WebGPU training reliability and llama.cpp-compatible checkpoints. With only tiny-model scope
- 08-18 13:28alert_routeThe public release is established and technically relevant, but its practical value depends on reproducing WebGPU training reliability and llama.cpp-compatible checkpoints. With only tiny-model scope
- 08-18 13:28groundA reliable direct-to-GGUF trainer would extend Scott’s existing local-model infrastructure and synthetic training-data work from inference/data preparation into portable model creation, while supporti
- 08-18 13:25promote_anchororigin walk conf 0.98
- 08-18 13:23createThe author reports a concrete open-source trainer and 95M checkpoint enabling a novel direct-to-GGUF local training workflow.