Skip to content

Navigation Menu

Appearance settings

View all features
- BY COMPANY SIZE
  Enterprises
  Small and medium teams
  Startups
  Nonprofits
- BY USE CASE
  App Modernization
  DevSecOps
  DevOps
  CI/CD
  View all use cases
- BY INDUSTRY
  Healthcare
  Financial services
  Manufacturing
  Government
  View all industries
View all solutions
- EXPLORE BY TOPIC
  AI
  Software Development
  DevOps
  Security
  View all topics
- EXPLORE BY TYPE
  Customer stories
  Events & webinars
  Ebooks & reports
  Business insights
  GitHub Skills
- SUPPORT & SERVICES
  Documentation
  Customer support
  Community forum
  Trust center
  Partners
View all resources
- COMMUNITY
  GitHub SponsorsFund open source developers
- PROGRAMS
  Security Lab
  Maintainer Community
  GitHub Stars
  Archive Program
- REPOSITORIES
  Topics
  Trending
  Collections
- ENTERPRISE SOLUTIONS
  Enterprise platformAI-powered developer platform
- AVAILABLE ADD-ONS
  GitHub Advanced SecurityEnterprise-grade security features
  Copilot for BusinessEnterprise-grade AI features
  Premium SupportEnterprise-grade 24/7 support
Pricing

Appearance settings

You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session. You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

interscript / rababa Public

Notifications You must be signed in to change notification settings
Fork 4
Star 13

Code
Issues 11
Pull requests 1
Actions
Projects
Security and quality
Insights

Additional navigation options

Code
Issues
Pull requests
Actions
Projects
Security and quality
Insights

Arabic + Hebrew SOTA stack, SadeedDiac-25 benchmark win, papers - #55

#55

Merged

ronaldtse merged 74 commits into

maininterscript/rababa:mainfrom

sota-sprint-arabicinterscript/rababa:sota-sprint-arabicCopy head branch name to clipboard

Aug 20, 2026

Conversation Commits74 (74)Checks Files changed

Merged

Arabic + Hebrew SOTA stack, SadeedDiac-25 benchmark win, papers#55
ronaldtse merged 74 commits into
mainfrom
sota-sprint-arabic

Commits

Commits on Aug 20, 2026

feat(arabic): Day 1 SOTA sprint — modern encoder + Muon + trie decoder
ronaldtse
committed
c8b56e2
feat(modal): wire huggingface secret into fetch_data for Sadeed HF download
ronaldtse
committed
92c283b
fix: MuonAdamWHybrid scheduler/GradScaler + server-side sota_pipeline
ronaldtse
committed
cbba8a6
fix(modal): cap sota_pipeline timeout at Modal max 24h
ronaldtse
committed
ac85751
fix(modal): commit datasets volume after fetch_data so pretrain can see corpus
ronaldtse
committed
ae0af29
feat(scripts): status.py — pull all 5 progress signals into one report
ronaldtse
committed
19ac5a2
fix(scripts): parse modal app list table correctly
ronaldtse
committed
e4c9b10
feat(hebrew): apply DS4/K3 modern stack — mHC + AttnRes + Muon + max_len 512
ronaldtse
committed
e7d364a
fix(modal): --force wipes old run-001 dirs and old artifacts so resume won't trip on stale format
ronaldtse
committed
9de9272
fix(modal): commit volume after force-wipe so next container sees clean state
ronaldtse
committed
6ed0be3
fix: shrink arabic_pro to 6L/512d (40M) + guard force-wipe by task
ronaldtse
committed
e50ebd0
fix(modal): key stage status by (task, stage) so parallel pipelines don't interfere
ronaldtse
committed
a9d1146
feat(scripts): status.py filters stages by task prefix for parallel pipelines
ronaldtse
committed
f5330de
feat: modern SOTA training stack for Arabic + Hebrew diacritization
ronaldtse
committed
80ecd8e
docs: publish results, papers, and TODO.publish execution log
ronaldtse
committed
d3ca746
docs: TODO.runtime-arch — training-to-usage pipeline work orders + agent prompt
ronaldtse
committed
45c82a2
docs(research): RL with verifiable rewards — GLM-5.3 playbook applied
ronaldtse
committed
3be1297
feat: Arabic ByT5-base trainer with per-save volume commits
ronaldtse
committed
3461072
feat: RAFT verifiable-reward RL + frozen private dev for Arabic
ronaldtse
committed
01d78a5
feat: haraqat error analyzer (word-final vs internal, confusion pairs)
ronaldtse
committed
0adbc85
feat: r3 domain-adaptation SFT on decontaminated Misraj corpus
ronaldtse
committed
14ff82e
fix: RAFT sampling survives preemption via incremental winner state
ronaldtse
committed
b80ea76
feat: r3 results, windowed eval, RAFT run-002 from r3
ronaldtse
committed
b9b1748
fix: windowed eval generation cap + haraqat projection
ronaldtse
committed
06d4815
fix: RAFT UnboundLocalError from nested json import
ronaldtse
committed
8937c8e
docs: final r3 windowed numbers in RESULTS + papers
ronaldtse
committed
aa4b315
feat: clean GLM (z.ai) eval on SadeedDiac-25
ronaldtse
committed
1badb46
results: GLM-5.2 verified on SadeedDiac-25 — 2.5060/1.5537
ronaldtse
committed
b5f5b4b
docs: GLM-5.2 verified frontier row in paper table
ronaldtse
committed
0f1979f
feat: r5 paragraph-context training (close the context gap)
ronaldtse
committed
7d3fb5c
feat: GRPO trainer — gold-reward RL with negative gradients
ronaldtse
committed
a79e1ee
perf: cache joined paragraph units on the volume
ronaldtse
committed
81cbbaa
docs: model manifest for the distillation agent
ronaldtse
committed
4568950
fix: RAFT/GRPO sized to survive the ~2h preemption cadence
ronaldtse
committed
2278c9e
perf: r5 checkpoints every 1000 steps
ronaldtse
committed
e5eb9be
fix: RAFT image missing accelerate — Trainer crash after sampling
ronaldtse
committed
feeec3c
fix: r5 OOM at 40%% — batch 3 + accum 10
ronaldtse
committed
a9d64ae
docs: distillation source-model usage prompt
ronaldtse
committed
152c682
fix: r5 OOM is deterministic — batch 2, unit cap 1450
ronaldtse
committed
2b84efe
feat: qalsadi morphological labeling for the aux-task lever
ronaldtse
committed
aa1f248
feat: GTPO entropy-weighted credit assignment in GRPO
ronaldtse
committed
ec762d2
docs: arXiv sweep Aug 2026 — GTPO applied, HomoRich/DIVRIT checked
ronaldtse
committed
78dec73
fix: two-tier morph labeling — exact tags + coarse fallback
ronaldtse
committed
c8a9695
fix: RAFT missing pyarabic; morph fallback on 'Not exists' analyses
ronaldtse
committed
fdf4e6e
feat: multi-reference WikiNews eval (QCRI EMNLP 2025 protocol)
ronaldtse
committed
f209c92
fix: stem-mode scoring excluded from totals, not auto-credited
ronaldtse
committed
8e68429
docs: r3 WikiNews-2024 multi-ref verdict (19.99/12.60 WER/DER)
ronaldtse
committed
573f9d0
fix: r5 OOM at 9048 — batch 1 / accum 30 (same effective batch)
ronaldtse
committed
b127779
docs: RAFT run-002 closed — flat on benchmark (2.8515/2.8308 vs r3 2.8429)
ronaldtse
committed
7bb48ca
fix: expose eval args through local entrypoint
ronaldtse
committed
3b88160
docs: r5 paragraph-context verdict — 2.6775/1.5965, beats verified GLM-5.2
ronaldtse
committed
7f2ae2f
docs: r5 SOTA in paper tables + MODELS manifest
ronaldtse
committed
2a4c213
docs: umbrella Arabic headline 2.68/1.60 (r5)
ronaldtse
committed
7b37f8a
fix: GRPO OOM — halve seqs per forward, detach entropy weights
ronaldtse
committed
b6051fb
docs: r5 WikiNews cross-domain tradeoff (+0.53 WER)
ronaldtse
committed
21979c4
fix: GRPO OOM root cause — enable gradient checkpointing
ronaldtse
committed
f3bda48
fix: graded alignment-based der() — binary reward was dead
ronaldtse
committed
10ffe8a
fix: GRPO throughput — 400 steps, save 50, eval 100
ronaldtse
committed
99e56af
fix: GRPO attention-score OOM — 1x16 micro-batch + SDPA
ronaldtse
committed
2d7a5fd
perf: GRPO right-sized for A100-80GB throughput
ronaldtse
committed
3aead24
fix: materialize best/ before final eval on flat curves
ronaldtse
committed
48102da
fix: GRPO image missing pyarabic (same as RAFT bug)
ronaldtse
committed
cd56105
docs: GTPO-GRPO closed flat — third negative RL result
ronaldtse
committed
211e168
docs: full results audit, blog post, refreshed distill prompt
ronaldtse
committed
493bb55
r6: morphological aux-task multitask (TAG: prefix format, 4x upsample, init r5, 24h timeout)
ronaldtse
committed
fb7a1ba
Hebrew s45: phonikud knesset 1.5M weak-pretrain then gold fine-tune (s43 recipe); decontam + dedupe prep
ronaldtse
committed
d59cbd1
s45 stage1: batch 16 x accum 4 + expandable segments (batch-64 OOM at 512 tokens)
ronaldtse
committed
b2697ea
distill prompt: in-flight successors, Persian v5 rescore verdict, WikiNews OOD note
ronaldtse
committed
0da029c
distill prompt: explicit start-now order; Thai gated on scaleup600k verdict
ronaldtse
committed
b3ef831
Thai teacher verdict: scaleup600k verified 1.7260% PER (was 2.32); distill gate now <=2.6 PER
ronaldtse
committed
7cc627c
distill order: Thai GO (verdict landed); s45 restarted note
ronaldtse
committed
285de08
MODELS: Thai 1.7260% PER (scaleup600k) — new best
ronaldtse
committed
0f6d715
Hebrew s45 phonikud curriculum verified 16.58% DER (s43 17.46): new teacher of record
ronaldtse
committed
e6f4ab2
r6: save every 300 steps (preemptions were outliving the 1000-step save interval)
ronaldtse
committed
bef1538

Footer

© 2026 GitHub, Inc.

Footer navigation

Terms
Privacy
Security
Status
Community
Docs
Contact

You can’t perform that action at this time.