---
language:
- en
license: apache-2.0
size_categories:
- 10K<n<100K
task_categories:
- text-generation
pretty_name: Agent Planning SFT (50K)
tags:
- agent
- planning
- react
- chain-of-thought
- sft
- supervised-fine-tuning
- reasoning
- tool-use
- agentic
- synthetic
configs:
- config_name: default
  data_files:
  - split: train
    path: agent-planning-sft-50k.jsonl
---

# Agent Planning SFT (50K)

50,000 ShareGPT-format conversations demonstrating ReAct-style agent reasoning: explicit Thought -> Action -> Observation cycles leading to grounded Final Answers across research, coding, data analysis, troubleshooting, planning, writing, and estimation tasks.

## Motivation

Most LLM training data shows answers, not the reasoning process that produces them. Agentic systems need models that can:
- **Decompose** complex tasks into concrete sub-steps
- **Identify** what information they need before answering
- **Verify** their reasoning with explicit observations
- **Synthesize** multiple findings into a coherent response
- **Flag** uncertainty and ask clarifying questions

This dataset trains on the full cognitive process, not just the final output.

## Dataset Description

**50,000 conversations** across 7 task types and 13 domains:

### Task Type Distribution

| Task Type | Count | % |
|---|---|---|
| Coding | ~15,400 | 31% |
| Research | ~11,500 | 23% |
| Planning | ~7,700 | 15% |
| Data Analysis | ~3,900 | 8% |
| Writing | ~3,900 | 8% |
| Reasoning | ~3,900 | 8% |
| Troubleshooting | ~3,900 | 8% |

### Domain Coverage

Factual lookup, comparative analysis, implementation, debugging, architecture, SQL analysis, production incident response, project planning, email drafting, Fermi estimation, decision support, code review, and technical roadmapping.

## Format

Standard ShareGPT format with ReAct-style reasoning in the assistant turn:

```json
{
  "conversations": [
    {
      "from": "human",
      "value": "How many piano tuners are there in Chicago?"
    },
    {
      "from": "gpt",
      "value": "Thought: This is a Fermi estimation problem. I need to estimate from first principles...\n\nAction: Estimate Chicago population\nObservation: Chicago city proper ~2.7M people...\n\nAction: Estimate piano ownership rate\nObservation: ~2.5% of urban households own pianos...\n\nFinal Answer: 25-50 piano tuners in Chicago city, 75-150 in the metro area."
    }
  ],
  "metadata": {
    "task_type": "reasoning",
    "domain": "estimation"
  },
  "id": "abc123"
}
```

## ReAct Pattern Demonstrated

Every assistant response follows the **Thought -> Action -> Observation** cycle:

**Thought**: What do I need to figure out? What are the key sub-problems?

**Action**: A concrete step taken (search, calculate, analyze, compare, write)

**Observation**: What was learned from that action - factual, not conclusory

**Final Answer**: Synthesis of all observations into a direct, well-structured response

This pattern trains models to reason transparently before committing to conclusions.

## Task Categories in Detail

**Coding tasks:**
- Implementation: write a thread-safe token bucket rate limiter
- Debugging: diagnose async ordering issues, trace execution paths
- Architecture: design webhook processing for 10k events/minute
- Code review: identify SQL injection, resource leaks, antipatterns

**Research tasks:**
- Comparative analysis: PostgreSQL vs MongoDB database selection
- Decision support: self-hosted ML vs managed API evaluation
- Factual lookup: LLC vs LP business entity comparison

**Planning tasks:**
- Project planning: monolith to microservices migration strategy
- Technical roadmap: test coverage improvement from 23% to 80%

**Data Analysis:**
- SQL analysis: consecutive 7-day login streak query (gaps and islands)

**Troubleshooting:**
- Production incident: API latency spike diagnosis without deployment

**Writing:**
- Email drafting: scope reduction announcement to stakeholders

**Reasoning:**
- Estimation: Fermi problem (piano tuners in Chicago)

## Key Training Signals

**1. Explicit decomposition**: Problems are broken into numbered action steps before the final answer appears. Models learn to plan rather than pattern-match to the first plausible response.

**2. Action specificity**: Each action is concrete and targeted ("Check for SQL injection vulnerability" not "look at the code"). Observations confirm or disconfirm specific hypotheses.

**3. Calibrated confidence**: Observations include uncertainty. The model acknowledges when assumptions are rough and provides ranges rather than false precision.

**4. Synthesis quality**: Final answers integrate all observations into a structured, actionable response - not just a list of findings.

**5. Question-back pattern**: Complex decisions end with a clarifying question to refine the recommendation, demonstrating that good agents identify missing information.

## Use Cases

- SFT fine-tuning for agentic AI assistants
- Training models to show reasoning work (chain-of-thought)
- Building AI systems for complex multi-step task completion
- Research on planning and decomposition in language models
- Training models to use tools and synthesize multi-source information
- Developer assistant products requiring debugging and code review

## License

Apache 2.0