Downloads · 30 days
9
12% of all-time downloads
le-el/tpn-003-song-hk1-final-20260906-smallcheck
tpn-003-song-hk1-final-20260906-smallcheck is a machine learning model from le-el. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This is preparation metadata for a trained, quantized text-inference candidate. The final 575-row repeat and two CPU memory checks completed on 2026-09-06 at 16:29:50 UTC; the local artifact SHA256 was freshly reverif…
Downloads · 30 days
9
12% of all-time downloads
All-time downloads
78
Public
Repo size
6.9 GB
Likes
0
Public
Click a slice to open those files.
.gguf6.9 GB · 100%
From the Hugging Face model README
This is preparation metadata for a trained, quantized text-inference candidate. The final 575-row repeat and two CPU memory checks completed on 2026-09-06 at 16:29:50 UTC; the local artifact SHA256 was freshly reverified. Local validation and preparation are complete. The metadata has not been published and does not certify official scores, competition eligibility or chain submission.
tpn003-v14-checkpoint-800-q5.gguffe616cb4bad7bfaa0f452cfaf621d550e2d8fe1de916b8c175e449480cab7f18Q4_K_S, with Q5_K token embeddings and no importance matrix. The filename's q5 label does not mean all tensors are 5-bit.On the same 575-row local five-task screen, this exact quantized candidate measured 0.7544597631, versus 0.6767679770 for the SFT-only C1 comparison artifact. These are weighted local screen composites, not full-suite or official service scores. Repeating those same 575 rows checks execution reproducibility; it is not an independent holdout evaluation. This comparison does not establish a rank or reward.
The completed repeat measured 0.7554393315, approximately 0.098 percentage points above the previous measurement of the same candidate. No new training occurred, and this small run-to-run difference is not presented as a training improvement. Task scores and evidence hashes are recorded in provenance.json.
The 27,392-row full evaluation was intentionally discontinued in favor of a smaller final check. No complete full-suite composite is reported. Completed task results and partial outputs are retained separately; incomplete work is not represented as a completed suite.
An earlier exact-artifact local precheck passed tokenizer, derivative-provenance and repeated memory/time checks. Those retained provenance checks are reused; the new two-run check tested memory and time only. Both new CPU runs passed in approximately 11.50 and 11.70 seconds, with logged allocation estimates of 7.58 GiB and peak RSS of approximately 7.74 GiB. The parser's 2,104,934 KiB figure covers only selected logged buffers, not total physical RAM usage.
This does not establish the validator's runtime identity. RAM parser values can change with the CPU backend; neither file size nor a local memory measurement is a certified submission RAM claim. No submission max_memory or official score claims are set by this package.
The direct HF source for local training and adapter merging was Lemura Labs Gemma-4-12B-uncensored-bf16. The official TPN-003 base is separately pinned as the competition provenance reference; its GGUF bytes are not asserted to be the direct parent of the trained HF weights.
Training used source partitions from MMLU, HellaSwag, ARC-Challenge and GSM8K. The retained recipe records exclusion of test answers and reserved holdouts; TruthfulQA was evaluation-only. Decontamination used exact normalized-question/context matching, not semantic or fuzzy matching. This package contains no training examples, benchmark questions or answers. Exact upstream revisions, training settings and evidence hashes are recorded in provenance.json.
The model can generate incorrect, unsafe or inappropriate content. No multimodal support, upstream endorsement, official benchmark parity, competition eligibility, rank, reward or warranty is asserted.
LICENSE preserves the exact Apache License 2.0 text from the retained source-evidence package. Its inclusion does not resolve which upstream model or dataset terms apply. Consult the upstream terms and retain LICENSE, NOTICE and this artifact's provenance.json with redistributed copies.