Sparidae/sft-agent-collection

SFT Agent Collection

This dataset is a curated mixture of agentic / tool-use / terminal SFT data, all converted into a unified messages+tools format.

Dataset Summary

Dataset	Source	Raw Samples	Filtered Samples	Parquet Size	Avg Tokens
Dolci-Instruct-SFT-Tool-Use	allenai/Dolci-Instruct-SFT-Tool-Use	227,579	122,410	205 MB	3,265
Dolci-Think-SFT-7B	allenai/Dolci-Think-SFT-7B	2,253,604	2,253,600	24 GB	11,152
LiteCoder-Terminal-SFT	Lite-Coder/LiteCoder-Terminal-SFT	11,255	11,150	166 MB	30,606
Nemotron-SFT-Agentic-v2	nvidia/Nemotron-SFT-Agentic-v2	47,602	47,602	243 MB	6,930
Nemotron-Terminal-Corpus	nvidia/Nemotron-Terminal-Corpus	139,841	57,405	636 MB	16,010
Open-SWE-Traces	Open-SWE-Traces	207,489	156,543	5.5 GB	70,156
SWE-smith-trajectories	SWE-bench/SWE-smith-trajectories	26,076	25,953	590 MB	34,249
ToolMind	modelscope/ToolMind	368,611	296,107	561 MB	2,583
xlam-function-calling-60k	Salesforce/xlam-function-calling-60k	101,363	71,896	80 MB	1,267
Total		3,383,454	3,042,666	31.5 GB	13,096

• Filter threshold: samples with num_tokens > 128,000 were removed (6,880 samples dropped).
• Tokenizer: Qwen/Qwen2.5-7B was used for token counting.

Schema

Each row in the parquet files contains the following columns:

Column	Type	Description
id	string	Unique sample identifier
source	string	Original dataset source
tools	string	JSON-serialized list of tool definitions
messages	string	JSON-serialized conversation messages
num_tokens	int64	Total token count of the sample
meta	string	JSON-serialized original metadata

Nested fields (tools, messages, meta) are stored as JSON strings to keep a consistent schema across heterogeneous sources.

Message Format

Messages follow the standard chat format:

• system: system prompt, usually contains tool definitions
• user: user input
• assistant: model response, may include tool_calls
• tool: tool execution results, references tool_call_id

Example tool call structure:

{
  "role": "assistant",
  "content": "",
  "tool_calls": [
    {
      "id": "call_xxx",
      "type": "function",
      "function": {
        "name": "get_weather",
        "arguments": "{\"city\": \"Beijing\"}"
      }
    }
  ]
}

Loading Example

import pyarrow.parquet as pq
import json

table = pq.read_table("Dolci-Think-SFT-7B.parquet")
messages = json.loads(table.column("messages")[0].as_py())
tools = json.loads(table.column("tools")[0].as_py())

Processing Pipeline

1. Download: raw datasets were downloaded from Hugging Face mirrors where network was unstable.
2. Phase 1 EDA: small-sample analysis to understand schema, tool usage, and message patterns.
3. Phase 2 Conversion: each dataset was converted to the unified format with tools and messages.
4. Token counting: num_tokens computed with Qwen/Qwen2.5-7B tokenizer.
5. Filtering: removed samples exceeding 128k tokens to fit common context windows.
6. Parquet compression: jsonl files were serialized to parquet with zstd compression for efficient storage.

Notes

• SWE-smith-trajectories raw token statistics count conversation messages only; patch fields (git diffs) are excluded because some exceed 100k tokens and are not part of the model input sequence.
• Open-SWE-Traces contributed the majority of dropped long-context samples due to its trajectory format.
• Some malformed JSON lines in Nemotron-SFT-Agentic-v2 were skipped during conversion.