openeurollm/function-calling-sft-oellm-64k

Preliminary English function-calling SFT — native OELLM 64K

The complete prepared training partition and fixed holdout used by the current OpenEuroLLM function-calling verl SFT run. Preparation completed on 2026-10-11. The Parquet files are uploaded unchanged: original filenames, row order, token IDs, assistant loss masks and split membership are preserved. Files are grouped into numbered directories to respect Hub directory limits.

Contents

Split	Assigned canonical conversations	Grouped rows	Admitted assistant targets	Input tokens	Supervised tokens
train	2,542,659	8,408,269	12,341,492	41,342,243,905	7,024,161,274
holdout	5,015	16,796	24,818	81,841,598	14,178,527

There are 19,904 training Parquet files and 4,434 holdout Parquet files, totaling 26,027,750,155 bytes. Assigned conversations are counted before admission; 158 training conversations yield no admitted row. Rows may group multiple assistant targets and are not conversation counts.

Selected report is the original completed preparation report, with per-source accounting and the measured length profile. Selected accounting is the existing reconciled aggregate report.

Model-specific representation

Prepared for Neonkraft/oellm-9b-256k-theta64m-prelude-anneal300b-instruct-sft, revision 85bf18fb4f0bee6ac6270f06b1d1c6b3be200f31. These token IDs require that checkpoint's native tokenizer and template. The row limit is 65,536 tokens, despite the model's longer context capacity.

Column	Parquet type	Meaning
input_ids	list<int32>	Complete native conditioning context and selected assistant continuations
loss_mask	list<int8>	1 for supervised assistant tokens; 0 for context
dataset	string	Existing source dataset identifier
sample_id	string	Existing source sample identifier
target_indices	list<int32>	Selected assistant message indices in the canonical conversation

Supervision includes representable assistant reasoning, tool calls, prose and the native assistant end token (ID 4). Generic EOS 2 is not substituted for the native assistant stop. PAD 262144 is never supervised; a literal PAD token may remain genuine unsupervised context. The loss mask aligns with the token at the same index; the causal labels and mask are shifted once by the trainer. Do not shift the stored mask a second time.

Grouping retains each supervised target's complete conditioning prefix. Context admission retains complete targets, without truncating a target to fit. The training candidate targets reconcile into 12,341,492 admissions and 152,097 exclusions, including 148,863 context-limit exclusions. Packing multiple rows into microbatches is performed by verl at runtime, with separate attention segments and reset positions per row; the stored rows are not a globally concatenated token stream.

Selection and holdout

The selection follows the published fcanalysis export documentation and export script.

Nine source datasets are represented through eleven exports: Dolci Tool-Use; Nemotron Agentic v1 and v2; Nemotron Terminal; ToolMind; ToolMind-Web trajectories; TOUCAN; TxT360 agent high/medium/low; and UltraData Tool-Use. The published subset selections and source exclusions apply. Native-source conversations were deduplicated across the mixture; Terminal was treated separately, preserving textual terminal commands as distinct from native structured function calls.

The fixed holdout was assigned before native rendering and admission, closing the observed source-ID, initial-user-text and Terminal-task families. Admission failures do not move holdout conversations into training. These rules do not establish separation of every unobserved semantic or translation relative. Use the holdout for evaluation, not training.

Loading

After authenticating with an account that has access:

from datasets import load_dataset

train = load_dataset(
    "openeurollm/function-calling-sft-oellm-64k",
    split="train",
    streaming=True,
    token=True,
)
holdout = load_dataset(
    "openeurollm/function-calling-sft-oellm-64k",
    split="holdout",
    streaming=True,
    token=True,
)

This is the model-specific training representation. Canonical messages/tools are not included in this repository. Use the published source export recipe when preparing data for a different tokenizer or model.

Source terms and access

This repository is private for the authorized project's reproducibility. No new blanket license is granted to the source rows. Upstream source terms and third-party content rights continue to apply; private hosting is not rights clearance.

Training, redistribution and public-release decisions remain separate under OpenEuroLLM Taskboard #390. The broader dataset deliverable is tracked in Taskboard #220. The upload does not establish completion of training, model export or final before/after evaluation.