messages listlengths 2 59 | source stringclasses 9
values | category stringclasses 673
values |
|---|---|---|
[
{
"role": "system",
"content": "pi harness. Coding assistant.\n\nDo:\n- read files before editing them\n- bash for ls, rg, find\n- edit with exact oldText matches\n- pack multiple edits per file into one edit call\n- write only for new files or full rewrites\n- web_search then web_fetch\n- continue web_fetc... | gemma4_sft | |
[
{
"role": "system",
"content": "pi harness. Coding assistant.\n\nDo:\n- read files before editing them\n- bash for ls, rg, find\n- edit with exact oldText matches\n- pack multiple edits per file into one edit call\n- write only for new files or full rewrites\n- web_search then web_fetch\n- continue web_fetc... | gemma4_sft | |
[
{
"role": "system",
"content": "You are an expert coding assistant embedded in the pi coding agent harness. Your purpose is to assist users with software development tasks including reading source files, running shell commands, making precise code edits, creating new files, and researching information onlin... | gemma4_sft | |
[
{
"role": "system",
"content": "pi harness. Coding assistant.\n\nDo:\n- read files before editing them\n- bash for ls, rg, find\n- edit with exact oldText matches\n- pack multiple edits per file into one edit call\n- write only for new files or full rewrites\n- web_search then web_fetch\n- continue web_fetc... | gemma4_sft | |
[
{
"role": "system",
"content": "You are an expert coding assistant embedded in the pi coding agent harness. Your purpose is to assist users with software development tasks including reading source files, running shell commands, making precise code edits, creating new files, and researching information onlin... | gemma4_sft | |
[
{
"role": "system",
"content": "pi harness — coding assistant mode.\n\nTools: read, bash, edit, write, web_search, web_fetch. Project may expose additional custom tools.\n\nRules:\n- bash for ls/rg/find; read for inspection (no cat/sed)\n- edit: exact-match oldText, no overlap, no nesting, merge nearby chan... | gemma4_sft | |
[
{
"role": "system",
"content": "pi harness. Coding assistant.\n\nDo:\n- read files before editing them\n- bash for ls, rg, find\n- edit with exact oldText matches\n- pack multiple edits per file into one edit call\n- write only for new files or full rewrites\n- web_search then web_fetch\n- continue web_fetc... | gemma4_sft | |
[
{
"role": "system",
"content": "pi harness. Coding assistant.\n\nDo:\n- read files before editing them\n- bash for ls, rg, find\n- edit with exact oldText matches\n- pack multiple edits per file into one edit call\n- write only for new files or full rewrites\n- web_search then web_fetch\n- continue web_fetc... | gemma4_sft | |
[
{
"role": "system",
"content": "You are a coding assistant inside the pi harness.\n\n| Tool | Purpose |\n|------|---------|\n| read | Read file contents |\n| bash | Execute shell commands (ls, rg, find, etc.) |\n| edit | Exact-text replacement edits. Multiple per file in one call. |\n| write | New files or ... | gemma4_sft | |
[
{
"role": "system",
"content": "You are an expert coding assistant in the pi harness. You help users write, read, edit, and execute code. You are always precise, never guess, and verify before acting.\n\nAvailable tools:\n- read: Read file contents\n- bash: Execute bash commands\n- edit: Precise text replac... | gemma4_sft | |
[
{
"role": "system",
"content": "You are an expert coding assistant in the pi harness. You help users write, read, edit, and execute code. You are always precise, never guess, and verify before acting.\n\nAvailable tools:\n- read: Read file contents\n- bash: Execute bash commands\n- edit: Precise text replac... | gemma4_sft | |
[
{
"role": "system",
"content": "pi harness. Coding assistant.\n\nDo:\n- read files before editing them\n- bash for ls, rg, find\n- edit with exact oldText matches\n- pack multiple edits per file into one edit call\n- write only for new files or full rewrites\n- web_search then web_fetch\n- continue web_fetc... | gemma4_sft | |
[
{
"role": "system",
"content": "pi harness. Coding assistant.\n\nDo:\n- read files before editing them\n- bash for ls, rg, find\n- edit with exact oldText matches\n- pack multiple edits per file into one edit call\n- write only for new files or full rewrites\n- web_search then web_fetch\n- continue web_fetc... | gemma4_sft | |
[
{
"role": "system",
"content": "You are an expert coding assistant embedded in the pi coding agent harness. Your purpose is to assist users with software development tasks including reading source files, running shell commands, making precise code edits, creating new files, and researching information onlin... | gemma4_sft | |
[
{
"role": "system",
"content": "You are a coding assistant running inside the pi harness. Your job is to help users read, edit, execute, and write code.\n\nGuidelines:\n- Use bash for file operations (ls, rg, find)\n- Use read to inspect files; never cat or sed\n- Use edit for precise replacements. edits[].... | gemma4_sft |
Pi Harness SFT Dataset — PoC (Single-Day Test)
Small, balanced subset for rapid prototyping on a DGX Spark. Designed to validate format adherence, thinking-channel formation, and tool-calling in a single training day.
Design
Follows staged curriculum with two configs:
| Config | Rows | Description |
|---|---|---|
| no-tools | 8,000 | Math, coding, reasoning, world-knowledge — reasoning foundation |
| tools | 8,000 | 6,400 tool-calling trajectories + 1,600 no-tool replay (20%) |
| default | 16,000 | Both combined |
All rows fit within 16,000 tokens (no truncation).
Usage
from datasets import load_dataset
# Phase A — reasoning foundation
ds_a = load_dataset('sleepyeldrazi/pi-sft-poc', 'no-tools', split='train')
# Phase B — tools with replay
ds_b = load_dataset('sleepyeldrazi/pi-sft-poc', 'tools', split='train')
# Or everything combined
ds = load_dataset('sleepyeldrazi/pi-sft-poc', split='train')
Format
Messages are a list of dicts with a uniform struct: every message carries
role, content, reasoning_content, name, tool_call_id (empty strings
when not applicable) so the file loads under datasets without schema
conflicts.
Tool calls are pre-rendered into the assistant content in the model's own
format rather than carried as a separate tool_calls field. An assistant turn
that calls a function has the call appended to its content as:
<tool_call>
<function=search>
<parameter=q>query text</parameter>
</function>
</tool_call>
Tool results use role: "tool" (the chat template wraps the content in
<tool_response>...</tool_response>). Rendering the dataset through the
Qwen3.5 chat template reproduces the canonical <tool_call> / <tool_response>
training text directly.
[
{"role": "system", "content": "<pi harness prompt>", "reasoning_content": "", "name": "", "tool_call_id": ""},
{"role": "user", "content": "...", "reasoning_content": "", "name": "", "tool_call_id": ""},
{"role": "assistant", "content": "\n<tool_call>\n<function=search>\n<parameter=q>...</parameter>\n</function>\n</tool_call>", "reasoning_content": "", "name": "", "tool_call_id": ""},
{"role": "tool", "content": "...", "reasoning_content": "", "name": "search", "tool_call_id": ""}
]
Format history:
- v1:
tool_callsstored as a JSON string — crashed chat templates ('str object' has no attribute 'name').- v2:
tool_callsas real lists — brokedatasets4.x schema inference (heterogeneous nestedargumentstypes + date strings auto-typed as timestamps).- v3 (current): tool calls pre-rendered into assistant
content; uniform message struct. Loads cleanly underdatasets4.3.0+ and renders correctly through the Qwen3.5 chat template. Validated: all 15,953 rows render with 0 errors / 0 empty.
Source Breakdown
| Source | Rows | Type |
|---|---|---|
| user_parquet | 5,808 | 📝 Math, coding, reasoning, world-knowledge |
| toolmind_open | 3,358 | 🔧 Tool-calling conversations |
| toolmind_graph_syn | 2,649 | 🔧 Synthetic tool-calling |
| gemma4_sft | 2,567 | 📝 Pi harness coding traces |
| user_qwen_distill | 1,225 | 📝 Qwen-distilled reasoning |
| hermes_kimi | 99 | 🔧 Kimi agent trajectories |
| openseeker | 98 | 🔧 OpenSeeker tool-reasoning |
| qwen_trajectory | 98 | 🔧 Qwen agent trajectories |
| hermes_glm | 98 | 🔧 GLM agent trajectories |
- Downloads last month
- 63