CPT on 161B tokens of Ukrainian court decisions — interested in updates?

#20
by overthelex - opened

Hi Joel,

I'm Vladimir Ovcharov, CEO of LEX AI LLC and a doctoral candidate at the Institute of Cybernetics (NAS of Ukraine), working on domain-adapted LLMs for Ukrainian law.

We've just completed a data preparation pipeline for Continued Pretraining (CPT) of Qwen2.5 on the full Ukrainian court decision registry (EDRSR):

  • Source: 38.5M court decisions → 33.9M after dedup + quality filtering
  • Tokens: 161.42B (Qwen2.5 tokenizer, fertility 0.515 tok/char)
  • Pipeline: 8-stage (export → clean → dedup MD5 → filter → structure → tokenize → package → S3), completed in ~3.5h on 208-core node
  • Training data: 19.7M sequences × 8192, 1,232 shards, 601 GB
  • Hardware: 8×H100 SXM 80GB (NVIDIA Innovation Lab grant)
  • Training: currently running CPT on Qwen2.5-14B (baseline) and Qwen2.5-32B (main), DeepSpeed ZeRO-3

We plan to evaluate on LEXTREME (Ukrainian subset) as one of our primary benchmarks. Our earlier experiments with generic models topped out at ~48.5 on LEXTREME — we're expecting a significant jump with domain CPT.

Would you be interested in hearing about the results? Happy to share the evaluation numbers, the training curves, and potentially the model weights once training completes.

We also have several related papers in progress:

  • Statute retrieval benchmark (arXiv: 2605.17639)
  • Citation graph analysis of Ukrainian case law
  • Temporal drift in legal language

Best regards,
Vladimir Ovcharov
CEO, LEX AI LLC
https://legal.org.ua
https://github.com/overthelex

Yes sure, please let me know :)

Hi Joel, quick update on the CPT run:

Qwen2.5-14B CPT — 22.7% complete

Metric Value
Steps 2,160 / 9,536 (22.7%)
Tokens seen 2.26B / 10B
Loss 0.837 → 0.255 (−69.5%)
Speed ~45s/step on 8×H100
ETA 3.8 days (May 27)
LR 4.5e-5 (cosine, warmup done)

Training is stable — no OOM since the initial cache pressure at step 1. Loss curve is smooth and still declining. The model is logging to MLflow for full reproducibility.

Plan after 14B completes:

  1. Evaluate on LEXTREME Ukrainian subset (the three epoch configs from our PR)
  2. Launch Qwen2.5-32B CPT with the same data (target 15B tokens)
  3. Share model weights + evaluation numbers here

Will post the LEXTREME eval results as soon as the 14B run finishes.

Best,
Volodymyr

Sign up or log in to comment