Full-training production version of oellm-9b-128k-theta32m. Long-context extension of OpenEuroLLM baby_9b_dense (Qwen3 dense 9B), 4K → 128K via native ABF (RoPE θ-scaling). Same θ=32M config as the sweep arm but trained on 2B tokens (≈6×). Base model — not instruction-tuned. Multilingual.

Evaluation — verified, multilingual (this checkpoint)

Base-LM forced-choice NIAH, depth-stratified, at 128K, 12 European languages:

metric	result
Overall 128K	100% (341/341)
depth-0 (far start)	100% in every language (6/6 each)
Languages	cs de el en es fi fr hu pl pt sv uk — all 100%
4K / 16K / 64K	100%

Full training lifted the two languages that were 96% in the sweep arm (el, sv) to 100%.

Why θ=32M (the key finding)

Far-position (depth-0) retrieval failure at long context is a RoPE high-dimension OOD problem, not a data problem — fixed by scaling θ to the target length. Critical θ ≈ doubles per length-octave: 64K→8M, 128K→16M→32M (100%), 256K→~32M. θ ablation at 128K depth-0: 2M/5M→0%, 8M→0%, 16M→90%, 32M→100%. (cf. LongRoPE2, arXiv:2502.20082.)

How to reproduce

Base: OELLM baby_9b_dense (Qwen3 dense): 36 layers, hidden 4096, FFN 12288, 32 heads / 8 KV (GQA), kv-channels 128, qk-layernorm, RMSNorm, SwiGLU, untied embeddings, vocab 262144, openeurollm/tokenizer-256k, native 4K.

Recipe: staged native ABF, 4K→16K→32K→64K(θ=2M)→128K(θ=32M); this model = the 128K@θ=32M stage trained to 2B tokens from the 64K checkpoint (--finetune).

Training (Megatron-LM, LUMI 16× MI250X):

--rotary-base 32000000 --seq-length 131072 --max-position-embeddings 131072 --use-flash-attn
--tensor-model-parallel-size 8 --context-parallel-size 8 --sequence-parallel --use-distributed-optimizer
--micro-batch-size 1 --global-batch-size 64 --bf16 --train-iters ~238 (2B tokens)
--lr 1e-5 --min-lr 1e-6 --lr-decay-style cosine --weight-decay 0.1 --clip-grad 1.0
--recompute-activations --recompute-granularity selective --save-interval 100
--qk-layernorm --normalization RMSNorm --swiglu --group-query-attention --num-query-groups 8

ROCm 6.4.4 / PyTorch 2.9 / TE 2.4 / FA 2.8; ~490 tok/s/GPU at 128K.

Data: Jouni Luoma's length-biased multilingual long-context mix (finepdfs all langs + edu, dclm, hplt3, multisynth, nemotron, megamath, starcoder, pes2o, arxiv, wiki; tiered short/medium/long). Note: depth-0 is insensitive to the data mix — θ is the lever.

Reproduce the eval:

python scripts/eval_base_lm_niah.py --model <this-model> \
  --context-lengths 4096 16384 65536 131072 --depths 0.0 0.25 0.5 0.75 1.0 \
  --languages en de fr es nl pl sv fi cs it pt el hu uk da --trials 6

Code / write-ups: https://github.com/BirgerMoell/openeuro-longctx-datamix

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
m = AutoModelForCausalLM.from_pretrained("birgermoell/oellm-9b-128k-theta32m-v3",
                                         torch_dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained("birgermoell/oellm-9b-128k-theta32m-v3")

Caveats

Base model (no instruction tuning) — use as a completion model.
Evaluated with single-needle forced-choice NIAH; broader multi-task RULER not yet run.
Keep rope_theta=32000000, max_position_embeddings=131072 for 128K.