@morphicode_jp

🇯🇵 個人で LLM の量子化リカバリを研究してます お仕事・共同研究・DM お気軽にどうぞ Independent AI researcher · DM open

Joined May 2026
  • Tweets43
  • Following118
  • Followers38
9Media
🚨2-bit Gemma 4 31B recovered to near 4-bit baseline performance by changing only 8 bytes. Q2_K patched (~12.6GB) vs unpatched Q4_K_M baseline (~19GB): HellaSwag 70.36% > 63.70% (+6.66pt) Winogrande 66.69% > 65.04% (+1.65pt) ARC-Challenge 46.44% > 44.89% (+1.55pt) GSM8k 93.20% < 95.80% (-2.60pt) So the honest result: Q2_K patched beats Q4_K_M baseline on 3/4 benchmarks. GSM8k remains Q4-favored. Across all four, Q2_K patched is within 0.69pt of unpatched Q4_K_M baseline. Patch: layer_output_scale[25] *= 1.5 layer_output_scale[26] *= 1.5 Two float32 values. 8 bytes total. No training. No calibration. No inference overhead. Paper: doi.org/10.5281/zenodo.20362… Code: github.com/morphicode-jp/f32… HF Q2_K: huggingface.co/morphicode-jp… HF Q4_K_M: huggingface.co/morphicode-jp…
16
16
2
199
1,127,191
日本人として応援してるよ。
Kodama@kodama26985649
Aug 18
llm-jp-4 の 33b dense モデルを公開しました！ 計算資源の不足もあって去年の秋から訓練し始めてようやく公開です… 順当に 8b や 32b-a3b よりは賢いです llm-jp.nii.ac.jp/news/202608…

約332億パラメータのDense型LLM「LLM-jp-4 33B」モデルを公開 - LLM-jp

大規模言語モデル研究開発センター（LLMC）では、オープンかつ日本語に強い大規模言語モデルの開発を進めており、2026年4月には、Dense型の「LLM-jp-4 8B」モデルと、Mixture of Experts（M […]

llm-jp.nii.ac.jp
95
期待大だけど、ローカルLLMの無検閲モデルの価値が落ちるような気がする
92
「お前何ができんねん。Fable 5より性能ええんか?」と聞いた結果 ⬇ 「は? 急に本気で聞いてきやがったなクソガキ」
1
65
5/6 名前の話。 探索は、全層をしらみつぶしに調べる screening と 0.05 刻みの細かい倍率探索。針を少しずつずらして効く一点を探し当てる——文字どおり「ツボ」を突く作業でした。だから TSUBO。日本語がそのまま手法名になりました。正式名は needle-point gate patching。 家族: TSUBO-8 (31B, 6/7 公開)、TSUBO-44 (31B, 11 層)、今回の TSUBO-4。数字は書き換え byte 数です。
1
1
100
6/6 仕組みの仮説: パッチは層ごとの F32 出力ゲートを増幅し、量子化で弱まった信号を補っている可能性。機構は未確定です。 GGUF + パッチ script + 論文: 📄 doi.org/10.5281/zenodo.20362… 🛠 github.com/morphicode-jp/f32… 🤗 hf.co/morphicode-jp/gemma-4-… RT / star / 🤗 like 歓迎です。

F32 Patches for Gemma 4 31B: 8-byte Cross-Quant Recovery (paper v1) and L25/L26 Per-Layer Ablation...

Combined release covering two related findings on Gemma 4 31B-it.Paper v1 (released 2026-05-27, included as primary PDF): an 8-byte F32 patch (layer_output_scale at layers 25 and 26, each multiplied...

zenodo.org
1
78
Local LLMs are entering a new era. nitter.cf/i/status/2063941078739…
平井 陽大 (Akito Hirai)@morphicode_jp
Jun 8
🚨2-bit Gemma 4 31B recovered to near 4-bit baseline performance by changing only 8 bytes. Q2_K patched (~12.6GB) vs unpatched Q4_K_M baseline (~19GB): HellaSwag 70.36% > 63.70% (+6.66pt) Winogrande 66.69% > 65.04% (+1.65pt) ARC-Challenge 46.44% > 44.89% (+1.55pt) GSM8k 93.20% < 95.80% (-2.60pt) So the honest result: Q2_K patched beats Q4_K_M baseline on 3/4 benchmarks. GSM8k remains Q4-favored. Across all four, Q2_K patched is within 0.69pt of unpatched Q4_K_M baseline. Patch: layer_output_scale[25] *= 1.5 layer_output_scale[26] *= 1.5 Two float32 values. 8 bytes total. No training. No calibration. No inference overhead. Paper: doi.org/10.5281/zenodo.20362… Code: github.com/morphicode-jp/f32… HF Q2_K: huggingface.co/morphicode-jp… HF Q4_K_M: huggingface.co/morphicode-jp…
4
572
公開延期
1
89
1/5 1-bit jumps to 60%. 4-bit just beat Q8. Gemma 4 31B-it just proved low-bit LLMs are back — and stronger than ever. Remarkably: 1-bit (IQ1_M, 9.5 GB) on GSM8k → 24% → 60% (+36.0pt) The largest improvement across all 12 cells. And 4-bit patched beats Q8 BF16 baseline on every single benchmark. No training. No calibration. No inference overhead. The conventional wisdom on low-bit LLMs has been completely overturned.
🤖 Made with AI
1
1
1
7
292
more replies
4/5  The patch works at every bit-level. Across the board. 1-bit  GSM8k +36.0pt 2-bit  HellaSwag +11.21pt 4-bit  GSM8k +15.0pt  (87% — beats Q8 baseline by +17pt) 4-bit  HellaSwag +9.80pt  (73.50% — beats Q8 baseline by +10.17pt) 12 cells, all positive. (See image for the full matrix.) No training. No calibration. No inference overhead.
🤖 Made with AI
1
4
84
5/5 Full reveal tomorrow (Mon May 25) · 22:00 JST. • Paper (arXiv preprint) • 3 GGUFs (1-bit / 2-bit / 4-bit) on HuggingFace • Patch application script on GitHub All free, all open source. CC-BY 4.0. Solo work. The new era of low-bit LLMs begins. Follow @morphicode_jp 🔔 to catch the drop live. Or bookmark this thread. Ready?
4
65
Load more