schoggie/java-agentic-recall-en

---
license: apache-2.0
language:
- en
- code
task_categories:
- text-generation
tags:
- java
- agentic
- code-generation
- synthetic
size_categories:
- 1K<n<10K
---

# Java Agentic + Recall (English)

Synthetic training data for fine-tuning a Java-specialist agentic coding model with explicit long-context recall capability. Companion dataset to a Qwen3.6-35B-A3B QLoRA SFT pilot.

## Composition

| Split | Source | Rows |
|---|---|---|
| train | DeepSeek V4 Pro (synthetic agentic Java traces) | 3873 |
| train | Synthetic positional recall — short context (~26K tok) | 120 |
| train | Synthetic positional recall — long context (50K-180K tok) | 46 |
| **train total** | | **4039** |
| eval | Synthetic positional recall — short (held-out) | 40 |
| eval | Synthetic positional recall — long (held-out) | 11 |
| **eval total** | | **51** |

## Schema

ShareGPT-style. Each row:

```json
{
"messages": [
{"role": "user", "content": "<problem statement or recall prompt>"},
{"role": "assistant", "content": "<solution / recall answer>"}
],
"source": "deepseek-v4-pro" | "synthetic-recall-short" | "synthetic-recall-long",
"metadata": { ... (recall examples only — method name, token bucket, source corpora) }
}
```

## Generation methodology

**Instruction traces (`deepseek-v4-pro`).** ~169 unique Java problem seeds (refactoring, debugging, library upgrades, concurrency, modern Java patterns) prompted to DeepSeek V4 Pro with a system prompt that asks for an agentic trace in `<plan>...<edit>...<bash>...<final>` tagged format. Targets Spring Boot 3.x, Java 17-21, modern enterprise idioms.

**Positional recall (`synthetic-recall-*`).** Real Java files from multiple Apache 2.0 codebases are stitched into multi-file documents at four token buckets (26K, 50K, 80K, 120K, 180K). For each document, methods are extracted (biased toward middle/late positions to test long-range attention) and the model is asked to reproduce the first 20 lines of a named method's body verbatim. Tests true long-context recall rather than near-window pattern matching.

Source repos used for recall (all Apache 2.0):
- Apache Commons Lang 3.14.0
- Spring Boot 3.2.0
- Spring Framework 6.1.0
- Jackson Databind 2.16.0
- Netty 4.1.100.Final

## Important notices

**AI-generated content.** All instruction traces in this dataset were generated by DeepSeek V4 Pro. Per DeepSeek's Open Platform Terms of Service §8.1, this content should be treated as AI-generated and may contain errors or omissions. Downstream consumers should validate code before using in production.

**License.** Content released under Apache 2.0 (compatible with the source Java code's license). Dataset compilation released under CC-BY-4.0.

## Limitations

- Code quality of synthetic traces is uneven; no human curation pass.
- Recall set tests verbatim reproduction, not semantic understanding.
- English-only; problem statements assume familiarity with Spring ecosystem.
- The 180K-token recall examples may exceed your model's context window — filter on `metadata.bucket_tokens` if needed.

## Citation

If you use this dataset, please cite as:

```bibtex
@misc{java_agentic_recall_en_2026,
title = {Java Agentic + Recall (English)},
author = {schoggie},
year = {2026},
publisher = {HuggingFace},
url = {https://huggingface.co/datasets/schoggie/java-agentic-recall-en}
}
```