File size: 8,718 Bytes
2e8ddb5 f13aec4 98e62ab f13aec4 2e8ddb5 98e62ab f13aec4 98e62ab f13aec4 98e62ab f13aec4 98e62ab f13aec4 98e62ab f13aec4 98e62ab f13aec4 98e62ab f13aec4 98e62ab f13aec4 98e62ab f13aec4 98e62ab f13aec4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 | ---
pretty_name: Coding Trajectory Dataset Samples
tags:
- coding-trajectories
- agentic-coding
- coding-agents
- software-engineering
- tool-use
- sft
- rlhf
- preference-data
- human-corrections
- failure-recovery
- verified-outcomes
- generative-ai
- code-llm
- agent-evaluation
license: cc-by-4.0
---
# Coding Trajectory Dataset Samples
## Overview
The **Coding Trajectory Dataset Samples** collection provides real-world agentic coding sessions in which an AI coding agent works with a developer inside a live software environment. Each session is captured as a structured trajectory covering developer requests, agent reasoning, tool calls, command results, file edits, and final outcomes.
Trajectories are organized at the session, turn, and event level, and are enriched with outcome labels (such as accepted, committed, overwritten, reverted, test passed, and test failed), human-correction signals, and session-level resolution scores. Curated views make the data ready for supervised fine-tuning, failure-recovery learning, verified-outcome training, and human-correction and preference research.
These resources can support research and development in coding agents, code-focused Generative AI, agent evaluation, reinforcement learning from human feedback, and software engineering automation.
## Dataset Highlights
- **Dataset Name:** Coding Trajectory Dataset Samples
- **Asset Type:** Agentic coding session trajectories
- **File Format:** Parquet (four linked tables)
- **Granularity:** Session, turn, and tool-call event level
- **Agent Surfaces:** Terminal and desktop application coding agents, with main-agent and sub-agent sessions
- **Model Coverage:** Multiple model tiers (Opus, Sonnet, Haiku, and specialized tiers)
- **Content Includes:** Developer prompts, agent reasoning, tool inputs and results, file edits, test outcomes, commits, and human corrections
- **Primary AI Use Cases:** Coding agent SFT, failure-recovery training, verified-outcome training, human-correction and preference learning, and agent evaluation
## Dataset Features
- Complete multi-turn coding sessions rather than isolated prompt-response pairs.
- Tool-call level events covering shell commands, file reads, edits, writes, searches, sub-agent calls, and external tool integrations.
- Outcome labels attached to individual events, such as accepted, committed, overwritten, reverted, test passed, and test failed.
- Human-correction records with the rejected action, the preferred action, the correction reason, and a quality score.
- Session-level summaries, outcome status, and resolution scores.
- Agent reasoning attached to turns, with the reasoning source recorded.
- Tracking of abandoned branches, so recovery behavior can be studied.
- Structured judge labels for developer intent, such as confirmation, clarification, redirection, and correction.
- Personal and environment-specific identifiers scrubbed from the data.
## Data Structure
The samples are organized into four linked Parquet tables, joined by `episode_id`.
### 1. Episodes (`episodes.parquet`)
One row per session. Includes agent and client surface, session role (main or sub-agent), models used, capture source, session summary, session outcome, resolution score, curated view flags, label counts, and human corrections.
### 2. Turns (`turns.parquet`)
One row per developer-agent turn. Includes the user message and its kind (human, slash command, interrupt, queued message), model and effort level, agent reasoning, turn labels, judge labels, turn metadata (files touched, lines added and removed, duration), and the ordered steps taken.
### 3. Events (`events.parquet`)
One row per tool call. Includes tool name, event kind (command, edit), model, effort level, tool input, tool result, event labels, supporting evidence, reasoning, and a record of what was scrubbed.
### 4. Views (`views.parquet`)
Curated, task-ready records. Each row belongs to one of four views:
- **sft_ready:** Sessions suitable for supervised fine-tuning.
- **failure_recovery:** Sessions where the agent hits a failure and recovers.
- **verified_outcomes:** Sessions with verified results, such as passing tests or commits.
- **human_corrections:** Cases where the developer corrected the agent, with rejected and preferred actions.
## Label Taxonomy
Event and turn labels include:
- **Outcome signals:** accepted, committed, overwritten, reverted
- **Test signals:** test passed, test failed
- **Human signals:** human correction, human rejected, user confirm, user clarify, user nudge, user redirect
- **Agent behavior:** agent asked, agent revised prior work, agent self-fixed
- **Turn type:** no edit, outcome accepted, outcome overwritten, tests passed, tests failed
## Session Environment Labels
Sessions are classified by environment and quality-assurance label:
- **human_like:** Sessions resembling real developer workflows.
- **no_cwd:** Sessions without a working directory.
- **bare-container:** Sessions run in minimal container environments.
- **testbed:** Sessions run in test or benchmark environments.
Sessions with conflicting labels are held out as a separate quality-assurance pool.
## AI Tasks Supported
The dataset may be useful for research and development involving:
- Coding Agent Supervised Fine-Tuning (SFT)
- Reinforcement Learning from Human Feedback (RLHF)
- Preference and Human-Correction Learning
- Failure Recovery and Self-Correction Training
- Verified-Outcome and Reward Modeling
- Tool-Use and Function-Calling Training
- Agent Evaluation and Benchmarking
- Long-Horizon Planning in Software Tasks
- Code Generation and Code Editing
- Developer Intent Understanding
## Potential Applications
- Training and fine-tuning AI coding assistants and autonomous coding agents.
- Building reward models from test results, commits, and accepted or overwritten edits.
- Studying how agents recover from failed commands and failing tests.
- Learning from developer corrections to align agents with real developer expectations.
- Evaluating agent behavior across models, effort levels, and environments.
- Analyzing real developer-agent collaboration patterns.
- Building debugging, refactoring, and code review automation.
- Researching tool-use behavior in agentic workflows.
## Industries and Research Areas
Potential areas of application include:
- Artificial Intelligence and Machine Learning
- Generative AI and Code LLMs
- Agentic AI
- Developer Tools and DevOps
- Software Engineering Automation
- Reinforcement Learning
- AI Safety and Evaluation
- Enterprise Software
- Education Technology (EdTech)
## Intended Audience
This dataset may be useful for:
- AI and machine learning researchers
- Generative AI and code LLM developers
- Coding agent and developer tool builders
- Reinforcement learning and alignment teams
- AI evaluation and benchmarking teams
- Enterprise AI teams
- Software engineering researchers
- EdTech companies
- Academic institutions
## Privacy and Data Handling
Trajectories are processed to remove personal and environment-specific identifiers. Scrubbed categories include local usernames, person names, GitHub handles, social profile IDs, personal sites, commit hashes, session and agent IDs, local proxy and temporary directory details, public IP addresses, and third-party names. Images are omitted from the samples.
## Advantages
- Real developer-agent sessions covering the full workflow from request to outcome.
- Rich outcome signals that support reward modeling and verification-based training.
- Human-correction records suited to preference learning.
- Failure and recovery behavior preserved, not filtered out.
- Multiple models and effort levels for comparative research.
- Task-ready curated views alongside full session detail.
- Personal identifiers scrubbed for safer research use.
## Summary
The **Coding Trajectory Dataset Samples** collection represents real agentic coding sessions captured at session, turn, and tool-call level, with outcome labels, human-correction signals, and curated views for SFT, failure recovery, verified outcomes, and human-correction learning. It can support research and development in coding agents, agent evaluation, RLHF, tool-use training, and software engineering automation.
**Data Creation**
Procured through formal agreements and generated in the ordinary course of business.
**Considerations**
This dataset is provided for research and educational purposes only. It contains only sample data. For access to the full dataset and enterprise licensing options, please visit our website [InfoBay.AI](https://infobay.ai/) or contact us directly:
- **Phone:** +91 9236398619
- **Email:** datareq@infobay.ai
|