sriq-ai/sriq-sft-v1.5

sriq-sft-v1.5

Superseded by [sriq-sft-v1.6](https://huggingface.co/datasets/sriq-ai/sriq-sft-v1.6). Only 0.9% of the verification steps in this version ever catch a mistake — the check runs on a trace rewritten from reasoning that was already correct, so it has nothing to find. A model trained here learns the shape of checking without the function. v1.6 fixes it; see Provenance.

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: a system turn, a human turn, then one GPT turn.
• —System turn: from is system; value contains either a short generic preamble or a long synthetic coding-harness preamble (see below).
• —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.

What v1.5 adds

v1.5 is v1.4 plus long system prompts that simulate agentic coding harnesses.

Models fine-tuned on v1.3/v1.4 hold Chinese reasoning under a short system prompt, but drift back to English reasoning once they are put behind a real agent harness — a multi-thousand token English system prompt full of tool descriptions, skill listings and project context. The reasoning language is conditioned on the prompt the model actually sees at inference, so the training data has to contain that prompt shape. v1.5 supplies it: 599 of the 1,424 rows carry a harness-style preamble, so Chinese reasoning is reinforced in exactly the context where it otherwise collapses.

Harness preambles are 913–15,976 tokens (mean 4,531, median 1,518) and are near-unique: 598 distinct harness names across 599 rows. Each is assembled from

• —a persona line naming a fictional harness — terminal coding agent, expert coding assistant, command line development assistant, software engineering assistant, or autonomous programming agent — inside a randomly named product;
• —a Guidelines: block (test after changes, never log secrets, prefer dedicated tools over bash, and so on) and a harness documentation file list;
• —a # Skills listing (all 599 rows) of real-world skill descriptions, themselves multilingual, in the format an agent harness injects them; 33 rows also carry a # Tools section, and 52.8% mention MCP;
• —a # Project Context block (62.4% of rows) quoting AGENTS.md / README.md / CONTRIBUTING.md / ARCHITECTURE.md fragments;
• —a trailing environment footer with the current date and working directory (all rows).

The user turn is a normal hard math or coding question, not a tool-call trace — the harness preamble is distractor context, and the target is that the model keeps reasoning in Chinese and answering in the question's language despite it. No row contains an executed tool call or a tool-result turn.

Also inherited from v1.4: reasoning chains close by checking the answer against the original constraints before the final answer.

Token statistics

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

Rows: 1424. All rows are difficulty: hard; topic is math (801) or coding (623). Mean reasoning length is 438 characters (467 on harness rows, 417 on the generic ones).

Languages (30, evenly present in both the harness and generic slices): 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-v1.5", 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

system, question = conversation[0]["value"], conversation[1]["value"]
response = conversation[2]["value"]
reasoning, answer = response.removeprefix("<think>").split("</think>", 1)
messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": question},
    {"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 system/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.

The harness rows are long: the longest system prompt alone is ~16k cl100k_base tokens. Set the sequence length to keep those rows whole — truncating the preamble removes the exact signal v1.5 exists to teach. If you must cap, drop the row rather than truncate it.

Note that the harness rows are appended at the end of this dataset, not interleaved. A tail split holds out almost exactly the rows a context cap would already exclude — shuffle before splitting.

Provenance

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-v1.5", split="train",
             revision="e7ee870519aa755a1ab285d8dc60f492c4e61824")

Release chain

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

What v1.5 left for v1.6 to fix. Three things, all measured on this corpus:

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