saidutta69's picture
docs: link contributor tool repo in contributing section
c846e32 verified
|
Raw
History Blame Contribute Delete
9.62 kB
metadata
license: mit
language:
  - en
pretty_name: Agentic Vibecoding Traces
tags:
  - agentic
  - agent-traces
  - tool-use
  - vibecoding
  - sft
  - distillation
  - multi-agent
  - multi-teacher
  - reasoning
  - chain-of-thought
  - anonymized
  - gated
task_categories:
  - text-generation
size_categories:
  - 1K<n<10K
extra_gated_prompt: >-
  This dataset contains anonymized personal agentic coding traces. By requesting
  access you agree to use it for research and model training only, not to
  attempt re-identification of the contributor, and not to redistribute the raw
  data.
extra_gated_fields:
  Name: text
  Email: text
  Organization or independent: text
  Intended use: text
  I agree to the data-use terms above: checkbox
configs:
  - config_name: premium
    default: true
    data_files:
      - split: train
        path: data/premium/train.parquet
      - split: validation
        path: data/premium/validation.parquet
      - split: test
        path: data/premium/test.parquet
  - config_name: standard
    data_files:
      - split: train
        path: data/standard/train.parquet
      - split: validation
        path: data/standard/validation.parquet
      - split: test
        path: data/standard/test.parquet
  - config_name: unfiltered
    data_files:
      - split: train
        path: data/unfiltered/train.parquet
      - split: validation
        path: data/unfiltered/validation.parquet
      - split: test
        path: data/unfiltered/test.parquet
  - config_name: opencode
    data_files:
      - split: train
        path: data/opencode/train.parquet
      - split: validation
        path: data/opencode/validation.parquet
      - split: test
        path: data/opencode/test.parquet
  - config_name: claude-code
    data_files:
      - split: train
        path: data/claude-code/train.parquet
      - split: validation
        path: data/claude-code/validation.parquet
      - split: test
        path: data/claude-code/test.parquet
  - config_name: grok-cli
    data_files:
      - split: train
        path: data/grok-cli/train.parquet
      - split: validation
        path: data/grok-cli/validation.parquet
      - split: test
        path: data/grok-cli/test.parquet
  - config_name: manicode-freebuff
    data_files:
      - split: train
        path: data/manicode-freebuff/train.parquet
      - split: validation
        path: data/manicode-freebuff/validation.parquet
      - split: test
        path: data/manicode-freebuff/test.parquet

🧠 Agentic Vibecoding Traces

RACER IS OP

3.5 years of real agentic coding sessions across 4 CLI agents and 25+ teacher models — fully anonymized, segmented per-task, with complete tool-call trajectories (bash commands + outputs, file edits) and chain-of-thought reasoning.

The culmination dataset: every "vibe coding" session, extracted from local agent storage, scrubbed, and packaged for SFT.

Gated access. Access requests are reviewed manually. Data is anonymized (projects → proj-<hash>, identities/secrets redacted). Do not attempt re-identification or redistribution.

📊 Dataset Overview

Property Value
Premium Segments 5,266 (default config — completed, deduped, scored ≥70)
Standard 5,838 (score ≥50)
Unfiltered 5,928 (everything, for RL/negative mining)
Duplicates Removed 187 exact task→solution pairs
Source Agents 4 (opencode, Claude Code, Grok CLI, Manicode/Freebuff)
Teacher Models 25+ (model column on every row)
Tool Calls ~52,000 (with inputs AND outputs)
Reasoning Text ~41M chars of CoT

✨ Quality Tiers

Quality > Ease of Access > Quantity

Config Rows Gate
premium (default) 5,266 score ≥70 + final answer present + no refusals + no tool-call loops
standard 5,838 score ≥50
unfiltered 5,928 everything — use for RL / negative samples / analysis

Scoring rubric (0-100): completed trajectory (+25), real tool work (+10), tool error rate (+15/+7), no repetition loops (+10), sane length (+5), refusal (-40), dead-end (-30).

🔌 Source Agents

Agent Config Premium Rows
opencode opencode ~4,500
Manicode / Freebuff manicode-freebuff ~1,000
Claude Code claude-code ~47
Grok CLI grok-cli ~10

Agent configs contain premium-tier rows only.

🤖 Teacher Models — filter per model

Every row carries a model column. Pick any single teacher:

from datasets import load_dataset

ds = load_dataset("saidutta69/agentic-vibecoding-traces", "premium", split="train")

# Single-model SFT — filter the model column
deepseek = ds.filter(lambda x: x["model"] == "deepseek-v4-flash-free")
oxalpha  = ds.filter(lambda x: x["model"] == "x-preview-f-free")
opus     = ds.filter(lambda x: x["model"] == "claude-opus-4-8")

# Or an agent-specific config (premium tier)
cc = load_dataset("saidutta69/agentic-vibecoding-traces", "claude-code")

# RL / negative mining
raw = load_dataset("saidutta69/agentic-vibecoding-traces", "unfiltered")

Top teachers by segment count:

Model Segments
deepseek-v4-flash-free ~3,700
manicode/freebuff ~1,050
mimo-v2.5-free ~420
minimax-m2.5-free ~120
hy3-free ~110
claude-opus-4-8 ~60
x-preview-f-free (Ox Alpha) ~47
glm-5.2 ~22
... 17 more ...

🧹 Anonymization Pipeline

Privacy-first. Every row passed through:

Step Detail
Project names All directory slugs (+ stem variants) → proj-<sha1-8>
Home paths /Users/<user>/...~/ everywhere incl. nested JSON
Identity strings Real name, handles, orgs → [REDACTED]
Emails [EMAIL] (2,130 replaced)
Secrets API keys, tokens (sk-, ghp_, bearer…) → [REDACTED] (1,794+)
Hostnames [HOST]
Session titles Dropped entirely (leak project context)

Verified zero occurrences of any known project name, handle, or identity string in the final parquets.

📂 Data Fields

Column Type Description
id string Stable sample id
agent string Source agent (opencode / claude-code / grok-cli / manicode-freebuff)
model string Teacher model — filter on this
split string train / validation / test (hash-seeded deterministic)
messages string (JSON) OpenAI-format messages incl. tool_calls (args as JSON), role:"tool" results, reasoning_content CoT
project_id string Anonymized project hash
quality_score int 0-100 rubric score
has_final bool Trajectory concluded with a substantive answer
error_loop bool ≥4 consecutive identical tool calls detected
err_rate float Fraction of tool outputs containing errors
n_tool_calls int Tool calls in segment
n_assistant_turns int Assistant turns
total_chars int Approximate size

🎯 Usage

from datasets import load_dataset
import json

ds = load_dataset("saidutta69/agentic-vibecoding-traces", "all", split="train")

row = ds[0]
msgs = json.loads(row["messages"])
# [{"role":"user","content":"..."},
#  {"role":"assistant","content":"","reasoning_content":"...","tool_calls":[{"id","function":{"name","arguments"}}]},
#  {"role":"tool","content":"...output..."},
#  {"role":"assistant","content":"final answer"}]

Works with TRL / Axolotl / LLaMA-Factory tool-calling templates. Segments are self-contained task→solution trajectories ready for assistant-only loss masking.

🤝 Contributing Your Sessions

Want to contribute your own agentic coding traces to grow this dataset?

Run the open-source cleaner on your own machine first: github.com/instax-dutta/vibe-trace-cleaner — it extracts sessions from your coding agents (opencode, Claude Code, Grok CLI, Manicode/Freebuff), anonymizes everything locally (projects hashed, identities/secrets redacted), and produces a reviewable output.

Then email contact@sdad.pro with your cleaned file (or questions) and I'll walk you through anything the tool doesn't cover, plus run a second independent privacy check before publishing.

Your sessions go through the same privacy pipeline: project names hashed, identities/secrets/emails redacted, titles dropped — nothing identifying ships.

⚠️ Notes & Limitations

  • Traces are real interactive sessions: some user requests are casual/incomplete; tool outputs may include long file dumps (kept for fidelity).
  • Teacher quality varies by model tier (free-tier models dominate counts) — filter by model for consistent teacher quality.
  • Contains terminal outputs from macOS/Linux environments; environment-specific details are inherent to agentic data.
  • MIT license applies to this packaging; underlying tool outputs are machine-generated session artifacts.

📜 Citation

@misc{agentic-vibecoding-traces,
  author = {Sai Dutta Abhishek Dash},
  title = {Agentic Vibecoding Traces: 3.5 Years of Multi-Agent Coding Sessions},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/saidutta69/agentic-vibecoding-traces}}
}