post-hoc checkpoint EMA (IMU-1): son N checkpoints_350m weight-avg -> ema/ckpt.pt (faz5 kiyasi), saf-mantik 3/3

v1.5b faz8_orpo: referans-modelsiz ORPO loss (NLL(chosen)-beta*logsig(log_odds)), chat-encode, saf-mantik 3/3 dogrulandi

v1.5b faz8: ORPO tercih verisi ({prompt-chat,chosen,rejected}), chat_prompt faz7_rag birebir, distilasyon notu, smoke-dogrulandi

v1.5b faz7_rag: inference chat sablonu (system/user + asistan-uretim), SYS/user_msg/is_tr faz7_prep import (train=inference), weights_only

v1.5b faz7: RAG-SFT {messages} converter (system=direktif/user=baglam+soru/assistant=kisa cevap), kisa-extractive+abstain, faz6 sablonu tek-kaynak, smoke-dogrulandi

v1.5b faz6: --cap_src smoke bayragi + uctan-uca dogrulandi (4 loader gercek veri, multi-turn, encode_any uyumu)

v1.5b faz6: cok-turlu {messages} maskeleme (encode_messages/encode_any) + base checkpoints_350m namespace + max_len 2048 (context)

v1.5b faz6: cok-turlu {messages} maskeleme (encode_messages/encode_any) + base checkpoints_350m namespace + max_len 2048 (context)

v1.5b faz6: cok-turlu sohbet SFT uretici (OASST2+TC-SFT+aya+Magpie, insan-oncelikli, {messages} format)

v1.5b: faz3_train.py — faz5 checkpoints_350m namespace + torch.load weights_only=False (resume/eval hardening)

v1.5b: faz5_eval.py — faz5 checkpoints_350m namespace + torch.load weights_only=False (resume/eval hardening)

v1.5b: emergency-save waits for async HF push (disconnect hardening)

v1.5b: 350M ckpt ayrı namespace (checkpoints_350m) — v1.0 177M ckpt ile resume çakışması fix

v1.5b: MIXES preset-aware (350m EN47/TR30 tc100b/code13/math10) + ShardStream mix param

decontam 13-gram index (eval suite, v1.5b tokenize icin)

faz1_01: --n_shards/--shard_index paralelize (ds.shard, ~8x)

faz1_01: --target_tokens bilimsel gosterim kabul (5e7/7.5e9)

faz3_train: fix print satiri args.n_layers->cfg[n_layers] (preset ile None oluyordu)

faz3_train: --preset 350m (cfg parametrize; 177m=177.1M v1-birebir, 350m=348.6M dogrulandi); --attn_every/--d_model vb override

faz1_01: tr_tc100b kaynagi (TC-100B ~103B TR, data_dir=pretrain) — v1.5b ana TR; 4-tuple data_dir destegi

faz3_train: Muon optimizer entegrasyonu (--muon: 2D-Linear Muon + embed/norm AdamW; cklu-opt state + WSD-carpan; geriye uyumlu)

faz9_recur: gen_khop ZINCIR gorevi (perm dongu-kisayolu kaldirildi; gercek k-hop)

lamba_cpu: cuda-guvenli (GQACPU._rope arange + Mamba3CPU.forward zeros device=)

faz9_recur: sys.path fix (HF symlink-cache yan-klasor import)

faz9_recur: --sweep (k zorluk taramasi, GAP-vs-k trendi = asil GO/NO-GO kaniti)

faz9_recur: --fast (hep-attention, Mamba token-loop yok, k-hop GO/NO-GO ~3-4dk GPU)

faz9_recur: Faz B harness (k-hop pointer-chase, fixed-vs-recurrent esit-param + test-zamani R olcekleme, device-esnek)

faz9_recur: recurrent-depth Faz A smoke (post-norm recursion kararliligi BIZIM bloklarla dogrulandi; gnorm patlamasi notu)

faz7_rag.py: --int8 dynamic quant (~2x kucuk bellek 708->325MB, ~1.2x hiz, ciktilar birebir)

lamba_cpu.py: --int8 dynamic quant (~2x kucuk bellek 708->325MB, ~1.2x hiz, ciktilar birebir)

faz7_rag.py: decode-cache (step/KV-cache) — CPU uretim O(L), ~6x hizli, full-recompute ile birebir

lamba_cpu.py: decode-cache (step/KV-cache) — CPU uretim O(L), ~6x hizli, full-recompute ile birebir

faz7_rag: --device cpu (LambaCPU saf-PyTorch yolu; fork import soft, CPU RAG yerelde calisiyor)

lamba_cpu: fix GatedMLP gate order (silu(a)*b -> y*silu(gate)) — CPU port now matches fork to 0.06-0.09 logit

lamba_cpu: GQA _rope dtype fix (bf16 uyumu)

lamba_cpu: EXP_E=True default (Mamba mixer KALIBRE, maxdiff 0.14)

lamba_cpu: EXP_E bayragi (alpha base-e vs base-2 decay)

lamba_cpu: BETA_SHIFT bayragi (fwd trap_shifted) + 4-grid

lamba_cpu: beta alpha-cikar bayragi (fwd kernel beta=DT(1-trap))

lamba_cpu RoPE: angle=tanh(ang)*DT*pi + interleaved (fork referans)

lamba_cpu: kalibre bayraklari (ANG_SIGN/ANG_DT/ROPE_INTER) grid-search icin

lamba_cpu kalibre 1: angle birikim isareti negatif

v1.1 lamba_cpu.py: saf-PyTorch CPU iskelet (load+forward calisiyor, matematik kalibrasyon bekliyor)

model card: Related Mamba-3 models + hrsvrn acknowledgement

LAMBA release stage: LAMBA_Inference.ipynb

LAMBA release stage: LAMBA_V1_MODEL_CARD.md

LAMBA release stage: faz7_rag.py

faz8_dpo: --length_norm (uzunluk yanliligi cozumu)

Faz 8 DPO: faz8_dpo.py

Faz 8 DPO: faz8_prep_dpo.py