File size: 8,444 Bytes
f7cd3b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-4.0
task_categories:
  - text-generation
  - text-classification
  - reinforcement-learning
language:
  - en
  - code
tags:
  - synthetic
  - coding-agent
  - mcts
  - reasoning-traces
  - process-reward-model
  - rlhf
  - dpo
  - agentic-ai
  - tool-use
  - code-generation
  - llm-training
  - ucb
  - reward-modeling
pretty_name: Coding Agent MCTS Reasoning Trace Pack
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: coding_intel_sample.parquet
---

# Coding Agent MCTS Reasoning Trace Pack (Sample)

**A synthetic Monte Carlo Tree Search reasoning-trace dataset for autonomous coding agents.** Each row is a complete reasoning lifecycle — initial context analysis → draft exploration → test feedback → prune-or-anchor → final outcome — labeled with a reasoning phenotype (TEST_DRIVEN, HACKER, DEEP_THINK, SECURITY_FIRST, REFACTOR_HEAVY) and carrying UCB scores at every step and explicit rewards at terminal actions.

Built by [SolsticeAI](https://www.solsticestudio.ai/datasets) as a free sample of a larger commercial pack. 100% synthetic. No real code, no proprietary repos — task titles and descriptions are generic archetypes drawn from canonical library patterns.

## What is included

| File | Rows | Format | Purpose |
|---|---:|---|---|
| `coding_intel_sample.parquet` | 10,000 | Parquet | Columnar, typed, best for analytics and RL training |
| `coding_intel_sample.jsonl` | 10,000 | JSON Lines | Streaming / LLM training friendly |

**Source pack:** 2.5M-trace corpus  
**This sample:** 10,000 reasoning traces, stratified 2,000 per reasoning phenotype  
**Reasoning phenotypes (5):** `TEST_DRIVEN`, `HACKER`, `DEEP_THINK`, `SECURITY_FIRST`, `REFACTOR_HEAVY`  
**Task types (3):** `bugfix`, `feature`, `refactor` (~3,300 each)  
**Languages (4):** `python`, `rust`, `go`, `typescript` (~2,500 each)  
**Production impact tiers:** `LOW`, `MEDIUM`, `HIGH`, `CRITICAL` (~2,500 each)

## Record structure

Each record is one reasoning lifecycle with 7 top-level fields:

| Field | Type | Contents |
|---|---|---|
| `schema_version` | string | Pack schema version (`1.0.0-coding-intel-sample`) |
| `event` | struct | `task_id`, `task_type`, `language`, `title`, `description` |
| `risk_context` | struct | `test_coverage_baseline`, `cyclomatic_complexity`, `production_impact` |
| `agent_reasoning` | list<struct> | Ordered reasoning steps: `action` (`analyze_context`, `write_draft`, `run_tests`, `lethe_prune`, `prometheus_anchor`), `depth`, `ucb_score` (null at root / terminal), `reward` (populated on terminal actions only), `thought` (natural-language rationale) |
| `correlated_telemetry` | struct | `linter_warnings_initial`, `linter_warnings_final`, `test_runtime_ms`, `ci_status` |
| `execution_summary` | struct | `files_changed`, `lines_added`, `lines_removed`, `time_to_resolution_sec` |
| `genetic_optimizer_feedback` | struct | `final_reward`, `lethe_prunes_triggered`, `nodes_expanded`, `phenotype_used` |

See [SCHEMA.md](./SCHEMA.md) for the full nested field breakdown.

## Why this dataset is useful

Most public coding datasets (HumanEval, SWE-bench, MBPP) only give you the *final answer* and the task description. They don't capture the reasoning tree the agent walked through — the wrong paths, the prunes, the anchor points. This pack is shaped around what modern agent-training pipelines actually need:

- **Explicit exploration vs exploitation.** Traces include both successful and pruned branches — `lethe_prune` events with negative reward, `prometheus_anchor` events with positive reward. Roughly 30% of traces carry a failed exploration branch before reaching the golden timeline.
- **Reward signals embedded at every step.** UCB scores at each non-terminal step, explicit rewards at terminal actions — directly usable for RL, DPO, and process-reward-model training.
- **Phenotype labels on every trace.** Train a `SECURITY_FIRST` coder specifically; run phenotype-transfer studies; build strategy-aware evaluation harnesses.
- **Correlated telemetry.** Linter-warning deltas, test runtime, and CI status correlated to reasoning outcome — grounds the trace in observable signals.
- **Compact.** Parquet fits in 340 KB, JSONL in 12.5 MB — you can pull this into a notebook in seconds and iterate.

## Typical use cases

- MCTS-based coding agent architecture training
- Process reward model (PRM) training
- Reasoning-chain evaluation benchmarks
- Agent self-improvement via trace replay
- Strategy-conditional code-generation research
- Curriculum learning with task-difficulty ladders
- LLM fine-tuning on structured reasoning narratives
- Benchmarking UCB-based exploration policies

## Quick start

```python
import pandas as pd
import pyarrow.parquet as pq

df = pq.read_table("coding_intel_sample.parquet").to_pandas()

# Phenotype distribution (stratified balanced)
print(df["genetic_optimizer_feedback"].apply(lambda g: g["phenotype_used"]).value_counts())

# Average final reward by phenotype
df["pheno"] = df["genetic_optimizer_feedback"].apply(lambda g: g["phenotype_used"])
df["reward"] = df["genetic_optimizer_feedback"].apply(lambda g: g["final_reward"])
print(df.groupby("pheno")["reward"].mean().round(2))

# Prune rate by task type
df["task"] = df["event"].apply(lambda e: e["task_type"])
df["prunes"] = df["genetic_optimizer_feedback"].apply(lambda g: g["lethe_prunes_triggered"])
print(df.groupby("task")["prunes"].mean().round(2))

# Pull one full reasoning chain
row = df.iloc[0]
for step in row["agent_reasoning"]:
    print(f"  d={step['depth']:<2} {step['action']:<20} ucb={step['ucb_score']} reward={step['reward']}: {step['thought']}")
```

Streaming form:

```python
import json

with open("coding_intel_sample.jsonl") as f:
    for line in f:
        trace = json.loads(line)
        # one MCTS reasoning trace per line
```

## Notes and limitations

- **Reasoning traces use canned action templates rather than live-executed code.** This pack is designed for agent-architecture training, not end-to-end SWE-bench-style evaluation.
- **`ci_status` is `SUCCESS` for every row in this sample** — the production pack includes `FAILURE` / `FLAKY` / `TIMEOUT` variants; this free sample is restricted to golden-timeline anchored traces to keep a clean reward surface.
- **UCB scores at root nodes use positive infinity** (serialized as `"Infinity"` in JSONL), following the standard MCTS convention.
- Phenotype distribution is uniform; production licensing supports custom phenotype mixes.

## Responsible use

This dataset is intended for **agent-training, process-reward-model, and MCTS research**. It contains synthesized reasoning narratives and action templates — it does **not** contain real code, real commit history, or proprietary repository content. Models trained on this data will learn reasoning structure and phenotype-conditional behavior; downstream code-generation quality still depends on training with real-code supervision from appropriately licensed corpora.

## License

Released under **CC BY 4.0**. Use freely for research, agent prototyping, education, and commercial development with attribution.

## Get the full pack

This Hugging Face repo is a **10K-trace sample**. The production pack scales to 2.5M+ traces with wider CI-outcome distribution (FAILURE / FLAKY / TIMEOUT), additional languages (C++, Java, Kotlin, Swift, C#), AST-diff variants, tool-call graph traces, multi-turn user-interaction sequences, custom phenotype mixes, and buyer-specific variants.

**Self-serve (Stripe checkout):**
- [**Sample Scale tier — $5,000**](https://buy.stripe.com/7sY5kD2j85QTfSb5lfeEo03) — ~25K records, one subject, 72-hour delivery.

**Full pack + enterprise scope:**
- [www.solsticestudio.ai/datasets](https://www.solsticestudio.ai/datasets) — per-SKU pricing across Starter / Professional / Enterprise tiers, plus commercial licensing, custom generation, and buyer-specific variants.

**Procurement catalog:**
- [SolsticeAI Data Storefront](https://solsticeai.mydatastorefront.com) — available via Datarade / Monda.

## Citation

```bibtex
@dataset{solstice_coding_intel_pack_2026,
  title        = {Coding Agent MCTS Reasoning Trace Pack (Sample)},
  author       = {SolsticeAI},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/solsticestudioai/coding-intel-pack}
}
```