cyankiwi/calibration-medium

A 16-language, 6-domain calibration dataset: 12,800 conversations ≈ 268.7M tokens, prioritizing agentic / tool-use traces, then code, math, stem, and structured output.

Schema

Four columns:

column	type	description
conversations	list<struct>	the chat transcript, one struct per message
language	string	ISO 639-1 code (16 values, table below)
language_type	string	domain: agentic, code, math, stem, general, structured_output
source	string	origin HF dataset repo id

Each message struct carries all five fields (null where absent):

role               : "system" | "user" | "assistant" | "tool"
content            : message text (tool output for role="tool")
reasoning_content  : assistant thinking trace (null when absent)
id                 : on role="tool" messages, the id of the tool_call it answers
tool_calls         : list<struct{id, type, function: struct{name, arguments}}>
                     (arguments is a JSON-encoded string; non-null only on
                      assistant turns that call tools)

Every role="tool" message's id matches a tool_calls[].id from a preceding assistant turn in the same conversation (validated over all 198,052 tool calls / 181,173 tool messages).

Loading

import json
from datasets import load_dataset

ds = load_dataset("cyankiwi/calibration-medium", split="train")

def preprocess(example, tokenizer):
    conv = []
    for m in example["conversations"]:
        msg = {"role": m["role"], "content": m["content"]}
        if m["tool_calls"]:
            msg["tool_calls"] = [
                {"id": tc["id"], "type": tc["type"],
                 "function": {"name": tc["function"]["name"],
                              "arguments": json.loads(tc["function"]["arguments"])}}
                for tc in m["tool_calls"]
            ]
        if m["role"] == "tool" and m["id"]:
            msg["tool_call_id"] = m["id"]
        conv.append(msg)
    return tokenizer.apply_chat_template(
        conv, tokenize=True, max_length=1024, truncation=True, return_dict=True)

Stratified subsets come directly from the columns, e.g. ds.filter(lambda r: r["language_type"] == "agentic") or ds.filter(lambda r: r["language"] == "ru").

Composition

Totals: 12,800 samples, 268,659,229 tokens (counted with the Qwen3-8B tokenizer over all text fields — content, reasoning, tool-call arguments).

Languages

Sample shares are exact quotas; token shares skew English because full-length agentic traces are kept unsplit (mean ~35k tokens).

lang	samples	sample share	token share	mean tok	p50	p90	max
en	5,376	42.0%	71.6%	35,765	34,832	69,234	185,073
zh	1,408	11.0%	6.1%	11,581	7,533	25,712	109,803
es	768	6.0%	3.8%	13,175	9,484	27,089	114,495
de	768	6.0%	4.0%	14,102	10,312	29,145	100,536
ar	512	4.0%	0.5%	2,822	2,644	4,609	15,221
fr	512	4.0%	2.4%	12,829	9,453	26,879	84,841
hi	512	4.0%	1.9%	9,735	6,312	22,814	113,897
ja	512	4.0%	2.4%	12,773	9,228	27,126	116,969
ko	512	4.0%	2.0%	10,490	7,023	22,718	62,713
ru	512	4.0%	2.9%	15,185	3,662	81,191	142,452
id	384	3.0%	0.2%	1,170	849	1,806	32,409
pt	384	3.0%	1.4%	9,667	7,235	20,952	47,382
th	192	1.5%	0.04%	596	452	818	4,839
vi	192	1.5%	0.1%	1,549	1,373	2,189	6,236
it	128	1.0%	0.6%	13,584	6,997	38,443	70,109
tr	128	1.0%	0.1%	1,088	722	2,061	8,571

Domains (language_type)

domain	samples	share	soft target	token share
agentic	3,680	28.7%	28.0%	64.0%
code	2,688	21.0%	24.0%	13.8%
math	2,062	16.1%	16.0%	13.1%
stem	1,856	14.5%	14.0%	2.9%
general	1,746	13.6%	12.0%	5.2%
structured_output	768	6.0%	6.0%	1.1%

Language quotas are hard; domain shares are soft targets. Code sits ~3pt under target (and general ~1.6pt over) because the floor-tier languages (th, vi, tr, id) only have general-domain sources.

Sources

samples	source	languages	domain(s)
3,520	nvidia/Nemotron-SFT-Multilingual-v1	de, es, fr, it, ja, zh	code, math, stem
1,760	nvidia/Open-SWE-Traces	en	agentic
1,600	nvidia/Nemotron-SFT-SWE-v3	en	agentic
1,568	nvidia/Nemotron-SFT-Multilingual-v2	hi, ja, ko, pt	code, math, stem
1,376	openbmb/UltraData-SFT-2605	en, zh	code, math, general, structured_output
1,152	allenai/Dolci-Instruct-SFT	ar, id, th, tr, vi	general
512	SupraLabs/agentic-reasoning-trace-summaries-40k	en	structured_output
416	Jackrong/Kimi-K2.5-Reasoning-1M-Cleaned	ar, ru	stem
303	MauroPello/multilingual-reasoning-gym-sft	hi, it, ru	math
256	AletheiaResearch/GLM-5.2-Agent	en	agentic
144	HelioAI/Fable-5-Distill-Reasoning-462x	en, ru	general
127	HelioAI/Claude-Fable-5-5500x	en, ru	math
64	Glint-Research/Fable-5-traces	en	agentic
2	kaist-ai/Multilingual-CoT-Collection	ru	general

Notes on specific sources:

• Open-SWE-Traces: only trajectories synthesized by MiniMax-M2.5 (minimax_m25 split), preferring resolved == 1 runs.
• GLM-5.2-Agent / Fable-5-traces: event-log agent sessions reconstructed into linear conversations (parentId chains; thinking blocks → reasoning_content, tool calls/results correlated by id).
• Kimi-K2.5: <think>…</think> blocks extracted into reasoning_content; ar/ru selected by script detection from the MultilingualSTEM subset.
• Fable-5-Distill-Reasoning-462x: the source has no answer field; the trace's closing conclusion was split out of the thinking trace into assistant content (marker-based with trailing-block fallback).
• Dolci-Instruct-SFT: ar/id/th/tr/vi selected by language ID (script ranges for ar/th; fastText lid.176 for id/tr/vi, with Indonesian accepting combined id+ms mass). None of the selected rows carry tool calls.

Construction

• Selection: streamed sampling (seed 42) with a minimum length of 1,024 tokens per conversation; full conversations kept unsplit (no truncation at build time). Global exact dedupe over message content.
• Backfill: when a language×domain×source bucket underfilled (source too small or too short), it was filled from same-language surplus in domain- priority order (agentic > code > math > stem > general > structured_output), then from the longest below-threshold rows. Below-1,024-token fills: id 217, th 179, tr 94, ru 31 — long rows in those languages are scarce in the available sources (e.g. only 1 of 60,570 scanned kaist ru rows passed the bar, hence its 2-sample contribution).
• Full per-bucket accounting (requested vs collected vs backfilled, token percentiles) is in composition_report.md / composition_report.json.

Caveats

• Token-share skew: English carries 71.6% of tokens despite 42% of samples — agentic SWE traces are long and kept whole. If samples are truncated to ~1k tokens (typical), effective language balance follows the sample shares, not token shares.
• Floor-tier lengths: th/id/tr samples are mostly below 1,024 tokens (th p50 = 452); they contribute genuine script/typology coverage but little long-context signal.
• Licensing: mixed — each row's source column points at its origin dataset; consult the source repos (CC-BY-4.0 / CC-BY-SA-4.0 for the Nemotron sets, Apache-2.0 for UltraData and SupraLabs, ODC-BY for Dolci, others vary).

Intended use

Calibration data. Not intended as an SFT corpus: sampling optimizes activation-statistics coverage (language × domain × rare-token mix), not answer quality.