sriq-ai/sriq-sft-v1.3

sriq-sft

Superseded. The current release is sriq-sft-v1.6. The next version after this one, sriq-sft-v1.4, adds a final logical verification step to the end of each chain of thought — v1.3 traces stop the moment a solution is reached. See Provenance for the full chain.

ShareGPT-compatible supervised fine-tuning conversations. Questions are generated in many languages. Assistant targets contain compressed Simplified Chinese reasoning and the original final answer. The final answer is not rewritten by the compression step.

• conversations: three turns — a system turn, a human turn, then one GPT turn.
• System turn: from is system; value is the assistant preamble.
• Human turn: from is human; value contains the question.
• GPT turn: from is gpt; value contains <think>Chinese reasoning</think> followed by the final answer.
• Extra columns: topic, difficulty, language (the requested question language).

There are no rejected responses in the SFT export.

Token statistics

Counts use tiktoken cl100k_base, not the training model's tokenizer.

Slice	Total tokens
Reasoning	725,302
Final answer	3,129,384
Reasoning + final answer, including think tags	3,868,471

Rows: 2703. Languages: Arabic, Bengali, Czech, Dutch, English, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Simplified Chinese, Spanish, Swahili, Swedish, Tamil, Thai, Traditional Chinese, Turkish, Ukrainian, Urdu, Vietnamese.

Load

from datasets import load_dataset
dataset = load_dataset("sriq-ai/sriq-sft-v1.3", split="train")
conversation = dataset[0]["conversations"]

Format for Qwen3.8 training

The export uses standard ShareGPT fields. Qwen3.8's native template expects reasoning in a separate field, so convert the record before applying that template:

from transformers import AutoTokenizer

turns = {turn["from"]: turn["value"] for turn in conversation}
response = turns["gpt"]
reasoning, answer = response.removeprefix("<think>").split("</think>", 1)
messages = [
    {"role": "system", "content": turns["system"]},
    {"role": "user", "content": turns["human"]},
    {"role": "assistant", "reasoning_content": reasoning, "content": answer.strip()},
]
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.8-27B")
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=False,
    preserve_thinking=True,
)

For other models, adapt the ShareGPT roles/fields to the trainer's expected conversation schema while retaining the complete GPT value. Do not place the same reasoning in both content and reasoning_content for Qwen's template.

For SFT, include the assistant's reasoning and final answer in the loss and mask user/prompt tokens. Check that the rendered training text retains the Chinese reasoning and that the assistant loss mask covers it. Use the target tokenizer to check lengths and keep repeated questions in the same train/evaluation split.

Provenance

Data revision	84f74b6a6d836e275771a9da896286d31d4704bf
Published	2026-09-17 09:04 UTC
Built with	sriq/sriq-sft, before that repository's history began — no commit corresponds
Superseded by	v1.4

The card above may change; the parquet at that revision will not. Pin it when training, so the corpus is the one that was measured:

load_dataset("sriq-ai/sriq-sft-v1.3", split="train",
             revision="84f74b6a6d836e275771a9da896286d31d4704bf")

Release chain

Each version exists to close a defect in the one before it.

Version	Rows	Data revision	What it introduced	The defect it left behind
v1.3	2,703	84f74b6	compressed Simplified Chinese reasoning, 30 languages	traces stop at the answer — 15.2% carry a 验 block, 0.3% end on an explicit judgement
v1.4	825	be3a4b6	a verification step closing the trace (95.4% carry 验)	Chinese reasoning reverts to English behind a long agent-harness prompt
v1.5	1,424	e7ee870	long synthetic harness preambles on 599 of 1,424 rows	the inherited check could only ever agree — 0.9% caught a mistake
v1.6	9,988	fa9e487	verification that catches a real planted mistake (19.5%), graded difficulty, a verification column	— current release

Pipeline and the full commit-level trail: git.nixre.dev/git/sriq/sriq-sft.