Commit History
-
Add paper title to speculative-decoding citation (arXiv:2211.17192)
Verified signaturee96ed6f
-
Add header image (was missing from this repo)
Verified signaturedbc5373 -
README: correct GGUF quant recipe + size accuracy
Verified signature- Describe the single shipped standard Q4_K_M (Q6_K-PLE) 5.34 GB GGUF instead of two distinct recipes — the two HF files are the same recipe (PLE tensor = Q6_K in both), verified by parsing the GGUF headers - Reframe the PLE-Q4_0 override as a 16 GB-hardware reproducibility recipe (~4.3 GB output), not a separately-shipped artifact - Fix residual 5.3 GB -> 5.34 GB and '10/10 benchmark' -> '10/10 project-held-out check' stragglersaa271ee
-
README: scope-lock MTP as future iteration
Verified signature- Rewrite MTP caveat: future-iteration framing, no measured-speedup claim - Consolidate citations to Google's six May 5 sources (blog + overview + HF collection + Kaggle + @GoogleGemma + arXiv:2211.17192) - Strip internal labels (run-N, variant-N, internal-plan refs, hardware specifics) - Preserve reviewer-runnable code locator (Section 14 / _RUN_MTP_DEMO=False) - Preserve variant-honest framing across the A4B + E4B familyd726395
-
Audit-driven refresh: P0 fixes (strip stale 8/8 benchmark, label W2C 2/3 as inferred-not-measured, remove insider notebook refs); P1 fixes (add Core Capabilities, Community Model, Technical Notes, Limitations, Citation sections to match A4B family structure); training details enriched with verified hyperparameters from finetune_run4.5; benchmark probes verbatim from inference_run7
Verified signaturea232a97 -
E4B card refresh: rename merged repo model-index to SolarHive-E4B-Merged; library_name peft->transformers; sync benchmark numbers to May 2026 final run (e4b 9/10, e4b-lora 10/10); expand tagline to name direct/Ollama/llama.cpp uses
Verified signaturecb9bf0f
-
Multi-Variant Deployment Validation (2026-02-05) + scaling-hypothesis docs
Verified signature- Multi-variant inference benchmark results (9-10/10 across 4 transformers variants) - When2Call validation: A4B 3/3 vs E4B 1/3 behavioral with paper citation - A4B-outperforms-E4B hypothesis with Google Gemma 4 docs verbatim quote - 5-tool registry + 2-message training/inference parity audit - E4B (d) probe matcher-passes-but-behavior-fails finding documented honestlyfdf66e6 -
Multi-Variant Deployment Validation (2026-02-05) + scaling-hypothesis docs
Verified signature- Multi-variant inference benchmark results (9-10/10 across 4 transformers variants) - When2Call validation: A4B 3/3 vs E4B 1/3 behavioral with paper citation - A4B-outperforms-E4B hypothesis with Google Gemma 4 docs verbatim quote - 5-tool registry + 2-message training/inference parity audit - E4B (d) probe matcher-passes-but-behavior-fails finding documented honestlyf934f11 -
Multi-Variant Deployment Validation (2026-02-05) + scaling-hypothesis docs
Verified signature- Multi-variant inference benchmark results (9-10/10 across 4 transformers variants) - When2Call validation: A4B 3/3 vs E4B 1/3 behavioral with paper citation - A4B-outperforms-E4B hypothesis with Google Gemma 4 docs verbatim quote - 5-tool registry + 2-message training/inference parity audit - E4B (d) probe matcher-passes-but-behavior-fails finding documented honestly9e78390 -
Multi-Variant Deployment Validation (2026-02-05) + scaling-hypothesis docs
Verified signature- Multi-variant inference benchmark results (9-10/10 across 4 transformers variants) - When2Call validation: A4B 3/3 vs E4B 1/3 behavioral with paper citation - A4B-outperforms-E4B hypothesis with Google Gemma 4 docs verbatim quote - 5-tool registry + 2-message training/inference parity audit - E4B (d) probe matcher-passes-but-behavior-fails finding documented honestlyfff1100
-
Restructure card to mirror solarhive-e4b-ollama (consistent layout: triple-role, How to Use code blocks, Companion Repos table, Fine-Tuning Architecture section)
Verified signaturedaa882f -
Initial upload of E4B v2 LoRA adapters from Run 4.5 Drive backup
Verified signature48abbb8 -
Correct E4B LoRA size estimate (~3 GB → ~200 MB) — actual adapter footprint at r=16 on E4B (8B base, 41M trainable params)
Verified signature0a7d1e4 -
Initial card: SolarHive E4B LoRA (v2 fine-tune, 0.9218 loss, 8/8 benchmark)
Verified signaturedc126ef -
initial commit
Verified signature52655e0