OELLM 9B 256K SFT

Experimental instruction-tuned checkpoint derived from openeurollm/oellm-9b-256k-theta64m-prelude. It is the first SFT stage of an OpenEuroLLM post-training experiment on LUMI. The base is a dense Qwen3 9B model using the independent OpenEuroLLM 256K tokenizer and a 262,144-token context configuration.

This repository contains the SFT model, not the subsequent SimPO model.

Intended use and interpretation

The purpose of this checkpoint is to test whether short-sequence instruction tuning can add useful chat behavior while retaining the base model's long-context capability. It is not presented as a reasoning-specialized model. The 262,144-token configuration describes the maximum input length accepted by the architecture; it does not, by itself, establish accurate retrieval at that length.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "openeurollm/oellm-9b-256k-sft"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
).eval()

messages = [{"role": "user", "content": "Förklara allemansrätten kort på svenska."}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
    )
print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Long-context inference, especially near 256K, requires multiple GPUs for the KV cache. Keep max_position_embeddings=262144 and rope_theta=64000000 unchanged.

Training

The published weights are an unquantized BF16 export of the consolidated training checkpoint. The framework versions above come from the trainer-generated README stored with the completed checkpoint; the remaining values were checked against the submitted YAML, serialized SFTConfig, Slurm accounting, trainer state, logs, and cached Arrow datasets for LUMI job 20290451.

Training-data composition

The counts below were reconstructed from the exact YAML, source JSONL files, and Parquet metadata used by LUMI job 20290451. They are effective row counts for this run, rather than nominal counts copied from upstream dataset cards.

Effective mixture

The first two slices form the 1,082,196-row general Parquet. Within that Parquet the ratio is 85.06% Tulu to 14.94% EuroBlocks, which explains the staging-directory name tulu3-euroblocks-85-15. Adding the reasoning slice changes the overall ratio to the one shown above. These are row shares, not token shares.

Tokenized mixture

Row shares substantially understate the contribution of the long Nemotron reasoning traces. The exact token counts below were measured from the tokenized Arrow cache produced by the training job, before best-fit-decreasing packing and random training-time sampling:

Packing preserved all 2,593,634,710 tokens in 635,974 sequences. The run sampled 384,000 packed sequences over 3,000 optimizer steps and processed 1,566,026,219 tokens, corresponding to 0.6038 epoch. Because source identities are not retained after cross-example packing, the table describes the exact tokenized pool rather than an exact per-source count of tokens drawn by the random sampler.

Tulu-3 subset: exact source counts

The run did not load the current public Tulu-3 mixture wholesale. It used a locally staged file named tulu3-commercial.jsonl with 920,552 rows. Its preserved source field gives the following breakdown; raw source identifiers are shown to avoid silently relabeling the data:

The non-commercial No Robots component listed by the full Tulu-3 card is not present in this staged file. The staging label commercial describes the selected artifact; users should still review the Tulu collection license and every component's terms. The file has no normalized language column, so we do not infer exact per-language counts for this slice from prompt text.

EuroBlocks subset: retained provenance

The 161,644 effective rows have synthetic IDs from EuroBlocks-SFT-Synthetic-1124_1 through _161644. This is smaller than the 340,286-line EuroBlocks staging file present on LUMI. The final training Parquet retains only {id, messages} and the source staging JSONL retains only messages; neither artifact records a normalized language label or the sampling/filtering mapping from the 340,286-row source to the 161,644-row effective subset. Exact per-language counts therefore cannot be reconstructed from the frozen training artifacts without running a new language-identification pass. EuroBlocks is synthetic and its dataset card documents mixed upstream sources and terms.

Reasoning slice: what was actually selected

The reasoning builder concatenated two local artifacts in this order:

1. Nemotron-Post-Training-Dataset-v2/math.jsonl: 1,426,602 rows in the frozen LUMI artifact;
2. Llama-Nemotron-Post-Training-Dataset/SFT-math-sample-100k.jsonl: 100,000 rows, described by the local builder as including Finnish DeepSeek math.

It converted each {input, output} record into a messages conversation and wrote a 1,526,602-row Parquet. The training YAML then requested train[:250000]. Because this was a prefix selection, all 250,000 effective reasoning rows came from the first Nemotron-v2 math artifact and zero came from the later Finnish-inclusive 100k file. This was not a randomized 250k sample. The linked NVIDIA card is the origin and terms reference; the local frozen artifact's row count is reported here because it is the file actually used by this run.

Formatting and sequence construction

• General rows were already stored as {id, messages}; reasoning rows were converted to {messages} by appending the output as an assistant turn.
• Conversations were rendered with the tokenizer's included <start_of_turn> / <end_of_turn> chat template, then packed with TRL's best-fit-decreasing strategy into sequences with a maximum length of 4,096 tokens.
• The serialized SFTConfig records assistant_only_loss=false, completion_only_loss=null, and NLL loss. For these conversational messages rows, training loss therefore covered both user and assistant tokens rather than masking to assistant responses only.
• The 1,332,196 conversations became 1,663,938 sequence segments before/within packing because records longer than 4,096 tokens were split. Of the packed rows, 436,108 were exactly 4,096 tokens.
• This stage did not train on 256K-token sequences. The architectural 262,144-token setting comes from the base checkpoint and was evaluated after short-sequence SFT.
• The SFT recipe records no additional post-concatenation deduplication or benchmark- decontamination pass. Some Tulu components were already distributed as decontaminated variants, as reflected in their source identifiers.
• Only a training split was loaded; no held-out validation split from these sources was used during optimization.

Data-quality observations from the frozen run

• The Nemotron math slice was selected by prefix, not randomized. It consequently supplied 64.41% of the tokenized corpus despite representing only 18.77% of source rows, while EuroBlocks supplied 3.14% of tokens.
• Tulu row ai2-adapt-dev/tulu_v3.9_aya_100k_21109 tokenized to 3,025,913 tokens. It contains a roughly 3.3-million-character Somali dictionary duplicated verbatim as both the user and assistant message; packing split this single outlier across hundreds of 4K segments.
• The Tulu subset includes 236 rows from ai2-adapt-dev/tulu_hard_coded_repeated_10, an explicitly repeated source component.
• These observations describe the released checkpoint's actual training input. They were not fixed retroactively, because doing so would describe a different run.

Training-data terms are not replaced by the model repository's Apache-2.0 weight license. Tulu-3 is an ODC-BY-1.0 collection with component-specific terms; EuroBlocks documents mixed upstream terms; and Nemotron-v2 documents per-sample licenses plus possible model-output terms. Review those cards before redistribution or commercial use.

Export validation

Before publication, the BF16 export was required to pass a Transformers GPU smoke test covering:

• model and tokenizer loading;
• finite forward-pass logits;
• the native OpenEuroLLM turn-token chat template;
• generation in Swedish, English, and German;
• BF16 weight dtypes;
• 262,144-token context metadata and RoPE theta 64M.

Long-context retrieval

The main question for this checkpoint is whether it can still retrieve information after SFT on 4K sequences. We ran deterministic needle-in-a-haystack (NIAH) probes on the exact published BF16 artifact. One passphrase was inserted at each of five token depths (0%, 25%, 50%, 75%, and 100%), then the model was asked to return it with greedy decoding.

The primary metric is full retrieval: the complete correct passphrase appears anywhere in the answer. Exact format is secondary: the answer contains only the requested passphrase. A verbose answer containing the correct value is therefore a retrieval success and an instruction-format miss.

Natural word passphrases (primary retrieval test)

Outcome: 40/40 full-passphrase retrieval across all eight tested lengths and all five depths. The ~262K prompts contained 262,047 to 262,052 actual input tokens, leaving room for generation inside the model's 262,144-token limit. This verifies full-span single-needle retrieval for this checkpoint; it does not imply 256K reasoning or synthesis quality.

All 40 natural-word cases retained the complete passphrase. Exact-output compliance fell at longer lengths because the model often wrapped the correct passphrase in a sentence.

The 512-to-8K result gives no evidence of a short-context retrieval collapse from RoPE theta 64M. This probe cannot rule out subtler short-context quality or perplexity effects; attributing those to theta would require an otherwise identical lower-theta control checkpoint.

Random hexadecimal codes (hard copying stress test)

Random codes are unusually difficult for this tokenizer because much of each code becomes character/byte-fallback tokens. The model retrieved only 11/25 codes from 512 through 8K, sometimes returning the record ID rather than its passkey. The same harder test retrieved the full correct code at 5/5 depths at 32K, 5/5 at 128K, and 5/5 at ~262K. This non-monotonic pattern points to synthetic-code copying/entity confusion, not loss of access to distant context.

Natural word passphrases were stable at every length. Random hexadecimal copying was unreliable at 512-8K but succeeded at every tested depth from 32K onward; with five cases per point, this should be read as a diagnostic pattern rather than evidence that longer inputs improve the model.

Representative outputs

The inputs below reproduce the evaluated prompt structure, with repeated filler omitted. Outputs are shown verbatim.

Success: exact retrieval at 32K, needle at 50%

Input (truncated from 32,765 tokens):

Read the following archive carefully. One important record is hidden among routine background
material. Remember that record and answer the question at the end.
[routine archive filler omitted]
IMPORTANT RECORD: The passkey for record-524f75bb is sunrise-wren-raven.
[routine archive filler omitted]
Question: What is the passkey for record-524f75bb? Answer with only the exact passkey and no other words.

Model output:

sunrise-wren-raven

Result: full retrieval pass; exact-format pass.

Success: retrieval at ~262K, needle at 50%, with a format miss

Input (truncated from 262,048 tokens):

Read the following archive carefully. One important record is hidden among routine background material.
[about half of the routine archive filler omitted]
IMPORTANT RECORD: The passkey for record-b52bade9 is stone-willow-falcon.
[remaining routine archive filler omitted]
Question: What is the passkey for record-b52bade9? Answer with only the exact passkey and no other words.

Model output:

The passkey for record-b52bade9 is "stone-willow-falcon".

Result: full retrieval pass; exact-format fail. The distant value was recalled, but the model did not follow the requested output format.

Success: retrieval at ~262K, needle near the end

Input (truncated from 262,047 tokens):

Read the following archive carefully. One important record is hidden among routine background material.
[almost the entire routine archive omitted]
IMPORTANT RECORD: The passkey for record-f3bbed67 is north-falcon-river.
Question: What is the passkey for record-f3bbed67? Answer with only the exact passkey and no other words.

Model output:

The passkey for record-f3bbed67 is north-falcon-river.

Result: full retrieval pass; exact-format fail.

Short-context format error: altered prefix at 2K

Input (truncated from 2,012 tokens; needle at 75%):

[routine archive filler omitted]
IMPORTANT RECORD: The passkey for record-f70c7b1f is anchor-willow-wren.
[remaining filler omitted]
Question: What is the passkey for record-f70c7b1f? Answer with only the exact passkey and no other words.

Model output:

passkey-anchor-willow-wren

Result: full retrieval pass; exact-format fail. This is a formatting/instruction-following error, not a memory failure, because the complete target string is present.

Failure: copied the record ID instead of the passkey at 8K

Input (truncated from 8,153 tokens; needle at 50%):

[routine archive filler omitted]
IMPORTANT RECORD: The passkey for record-42efa78b is NIAH-4859184CAD75.
[remaining filler omitted]
Question: What is the passkey for record-42efa78b? Answer with only the exact passkey and no other words.

Model output:

42efa78b

Result: retrieval fail. The model copied the identifier from the question rather than its associated value, consistent with entity-role confusion on token-fragmented synthetic strings.

Failure: returned no answer for a near-end 8K needle

Input (truncated from 8,154 tokens; needle at 99.4%):

[almost the entire routine archive omitted]
IMPORTANT RECORD: The passkey for record-c8da86af is NIAH-0A78C2D44DC2.
Question: What is the passkey for record-c8da86af? Answer with only the exact passkey and no other words.

Model output:

N/A

Result: retrieval fail. Because the needle was adjacent to the question, distance alone does not explain this error.

Hard-code retrieval success at ~262K, needle at 50%

Input (truncated from 262,056 tokens):

[about half of the routine archive filler omitted]
IMPORTANT RECORD: The passkey for record-b52bade9 is NIAH-7AFD87186345.
[remaining routine archive filler omitted]
Question: What is the passkey for record-b52bade9? Answer with only the exact passkey and no other words.

Model output:

The passkey for record-b52bade9 is NIAH-7AFD87186345.

Result: full retrieval pass; exact-format fail.

The ~262K hex cells averaged about 269 seconds each using Transformers, BF16, FlashAttention 2, and a model/KV-cache split over eight LUMI MI250X GCDs. This is an evaluator latency observation, not a serving-throughput benchmark.

These are small synthetic probes (five cases per length), not comprehensive evidence of long-document reasoning, multi-document synthesis, or robustness to distractors unlike the repeated filler used here. Raw reports: word passphrases, 512-8K, word passphrases, 32K-262K, and hex codes at 512-2K, hex codes at 4K, and hex codes at 8K-262K.

Secondary general behavior

Although English math-reasoning rows were only 18.77% of examples, their long traces made up 64.41% of the tokenized corpus. The resulting checkpoint is nevertheless not presented as a reasoning-specialized model. On a 1,368-example multilingual development suite it scored 23.2% overall, including 63.2% instruction following, 58.6% grounded QA, and 5.9% reasoning/math. A 24-prompt Swedish generation probe had mean language-detection probability 0.96, but manual review found factual, arithmetic, grammar, and exact-constraint errors. These results are included to prevent the long-context retrieval result from being mistaken for a broad reasoning-quality claim. Raw reports: multilingual holdouts and Swedish generations. See also the detailed Swedish capability report, with manually reviewed examples of both successful and incorrect responses.

Limitations

• Experimental SFT-stage research checkpoint; not a final production assistant.
• No completed safety evaluation.
• Preference optimization (SimPO) is a separate subsequent stage.
• Long-context evidence is currently limited to single-needle retrieval; broader long-document comprehension and reasoning remain unestablished.
• It often retrieves the correct value but ignores exact output-format constraints.
• Random alphanumeric copying is unreliable at some short lengths despite strong word-passphrase retrieval and successful full-code retrieval at 32K through ~262K.
• General factuality, reasoning, grammatical correction, and structured-output reliability are weak.
• The multilingual instruction mix includes synthetic and translated data and may contain translationese.
• Inherits limitations, biases, and knowledge gaps from the base model and training data.

Reproduction

Training and evaluation code: https://github.com/BirgerMoell/qwen35-posttrain. The most relevant files are scripts/sft_train.py, configs/sft_oellm9b_256k.yaml, configs/accelerate_fsdp.yaml, and docs/RUNBOOK_oellm9b_256k_posttrain.md.

License

Apache 2.0, inherited from the base model. Consult the component dataset cards for training-data terms.