Datasets:
id stringlengths 16 31 | repo stringclasses 54
values | kind stringclasses 7
values | teacher stringclasses 2
values | teacher_provider stringclasses 2
values | n_turns int64 4 120 | prompt_tokens int64 18.8k 7.15M | completion_tokens int64 288 104k | messages listlengths 9 242 | tools listlengths 8 8 |
|---|---|---|---|---|---|---|---|---|---|
arrow-018235b1a59d | arrow | single | Hy4 Flash | opencode-go | 17 | 197,504 | 3,878 | [
{
"role": "system",
"content": "You are an AI agent running in OpenCode, a coding agent harness. Help the user accomplish their goals using the tools you have available.\n\n# Harness\n- Responses are rendered as GitHub-flavored Markdown.\n- `<system-reminder>` blocks are harness instructions, not user-autho... | [
{
"type": "function",
"function": {
"name": "edit",
"description": "Edit the contents of a file by finding and replacing exact text. When editing text from Read output, preserve the exact indentation (tabs or spaces) and omit the line-number prefix, such as `1: `. Never include the prefix in old... |
arrow-065fe2210359 | arrow | multi | Qwen4 Max | nous | 20 | 346,511 | 6,710 | [
{
"role": "system",
"content": "You are an AI agent running in OpenCode, a coding agent harness. Help the user accomplish their goals using the tools you have available.\n\n# Harness\n- Responses are rendered as GitHub-flavored Markdown.\n- `<system-reminder>` blocks are harness instructions, not user-autho... | [
{
"type": "function",
"function": {
"name": "edit",
"description": "Edit the contents of a file by finding and replacing exact text. When editing text from Read output, preserve the exact indentation (tabs or spaces) and omit the line-number prefix, such as `1: `. Never include the prefix in old... |
arrow-0bb4d265e836 | arrow | implement | Qwen4 Max | nous | 18 | 326,893 | 7,011 | [{"role":"system","content":"You are an AI agent running in OpenCode, a coding agent harness. Help t(...TRUNCATED) | [{"type":"function","function":{"name":"edit","description":"Edit the contents of a file by finding (...TRUNCATED) |
arrow-1396b5603326 | arrow | single | Hy4 Flash | opencode-go | 10 | 74,580 | 1,172 | [{"role":"system","content":"You are an AI agent running in OpenCode, a coding agent harness. Help t(...TRUNCATED) | [{"type":"function","function":{"name":"edit","description":"Edit the contents of a file by finding (...TRUNCATED) |
arrow-1458a1b0cf46 | arrow | implement | Hy4 Flash | opencode-go | 6 | 36,123 | 1,013 | [{"role":"system","content":"You are an AI agent running in OpenCode, a coding agent harness. Help t(...TRUNCATED) | [{"type":"function","function":{"name":"edit","description":"Edit the contents of a file by finding (...TRUNCATED) |
arrow-153a2a675458 | arrow | implement | Hy4 Flash | opencode-go | 9 | 77,635 | 3,224 | [{"role":"system","content":"You are an AI agent running in OpenCode, a coding agent harness. Help t(...TRUNCATED) | [{"type":"function","function":{"name":"edit","description":"Edit the contents of a file by finding (...TRUNCATED) |
arrow-1a9672276bfd | arrow | test_only | Hy4 Flash | opencode-go | 10 | 77,742 | 1,796 | [{"role":"system","content":"You are an AI agent running in OpenCode, a coding agent harness. Help t(...TRUNCATED) | [{"type":"function","function":{"name":"edit","description":"Edit the contents of a file by finding (...TRUNCATED) |
arrow-1dad27ee7172 | arrow | single | Hy4 Flash | opencode-go | 7 | 40,593 | 790 | [{"role":"system","content":"You are an AI agent running in OpenCode, a coding agent harness. Help t(...TRUNCATED) | [{"type":"function","function":{"name":"edit","description":"Edit the contents of a file by finding (...TRUNCATED) |
arrow-23cb726c4f02 | arrow | single | Hy4 Flash | opencode-go | 14 | 110,156 | 1,746 | [{"role":"system","content":"You are an AI agent running in OpenCode, a coding agent harness. Help t(...TRUNCATED) | [{"type":"function","function":{"name":"edit","description":"Edit the contents of a file by finding (...TRUNCATED) |
arrow-2a3f56ca85ac | arrow | single | Hy4 Flash | opencode-go | 7 | 42,954 | 696 | [{"role":"system","content":"You are an AI agent running in OpenCode, a coding agent harness. Help t(...TRUNCATED) | [{"type":"function","function":{"name":"edit","description":"Edit the contents of a file by finding (...TRUNCATED) |
Qwen4Max-Hy4Flash-Distill
8,529 multi-turn agentic coding trajectories for supervised fine-tuning. Every trajectory was verified by running the repository's test suite, and only passing runs are included.
The trajectories were recorded from two teacher models working inside the OpenCode agent harness on real open-source Python repositories, with full tool use (read, edit, write, grep, bash, ...) and the teachers' reasoning:
| teacher | examples |
|---|---|
| Hy4 Flash (high reasoning) | 7,441 |
| Qwen4 Max (xhigh reasoning) | 1,088 |
Task kinds
| kind | examples | what the agent gets |
|---|---|---|
single |
2,120 | a GitHub-style bug report for one injected bug; fix it |
multi |
2,237 | a report covering several bugs across files; fix all |
implement |
2,269 | a function body was removed; implement it from the docstring / issue |
multi_impl |
243 | several stubbed functions to implement together |
tdd_fix |
461 | failing test output only; find and fix the cause |
test_only |
619 | a bug with no description beyond failing tests |
write_test |
580 | write tests that catch a described bug |
Bugs are AST mutations (negated conditions, swapped arguments, removed statements, flipped comparisons, ...) and LLM-written realistic rewrites, applied to lines that the repo's own tests cover. Issue texts were written by an LLM from the diff and are written not to reveal the fix. A task passes only if the previously failing tests pass and no previously passing test breaks.
Diversity: 54 repositories (arrow, attrs, click, jinja, lark, marshmallow, mpmath, networkx, pyparsing, sqlparse, textual, tomlkit, werkzeug, yapf, ...); the five largest repos together make up 15% of examples. On average 22 assistant turns per trajectory.
Format
JSON Lines, one trajectory per line:
{
"id": "arrow-018235b1a59d",
"repo": "arrow",
"kind": "single",
"teacher": "Hy4 Flash",
"teacher_provider": "opencode-go",
"n_turns": 17,
"prompt_tokens": 197504,
"completion_tokens": 3878,
"messages": [
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "assistant", "content": "", "reasoning_content": "...",
"tool_calls": [{"id": "...", "type": "function",
"function": {"name": "read", "arguments": {"filePath": "..."}}}]},
{"role": "tool", "tool_call_id": "...", "content": "..."},
...
],
"tools": [ ... OpenAI-style function schemas ... ]
}
messagesis in the OpenAI / Hugging Face chat format. Assistant turns carry the teacher's reasoning inreasoning_content; tool-callargumentsare JSON objects (not strings).toolsholds the exact tool schemas the teacher was given.prompt_tokens/completion_tokensare the teacher's usage summed over the trajectory.- The system prompt is OpenCode's default agent prompt without its "# Your Model" block (which normally names the serving model). Add your own model's identity there if you want it. Trajectories in which the model named itself were removed, so the data is not tied to any particular model identity.
- Trajectories are long: roughly 15% exceed 40k tokens (mostly
multiandimplement); filter by length for your context window.
Licensing and terms
The dataset card is Apache-2.0. Two things inside it come with their own terms, which you must follow too:
- Teacher outputs: assistant turns were generated by Hy4 Flash (Tencent) and Qwen4 Max (Alibaba). Check those models' terms of use for training on their outputs before using this data.
- Repository code: tool outputs contain source code from the 54 open-source repositories, each under its own license (mostly MIT, BSD and Apache-2.0).
Limitations
- Python only, single-repository tasks with injected bugs; real-world issues are messier.
- Only successful trajectories are kept, so the data shows no recovery from failed final attempts.
- Teacher reasoning is included as produced and can be long or wander.
- Downloads last month
- -