Datasets:

rafuke
/

freertos_esp32

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - freertos
  - esp32
  - esp-idf
  - qwen
  - unsloth
  - code
pretty_name: FreeRTOS ESP32
size_categories:
  - 1K<n<10K

FreeRTOS ESP32

Instructional ChatML dataset for fine-tuning language models (e.g. Qwen via Unsloth) as an expert ESP32 / ESP-IDF / FreeRTOS assistant.

Content is original instructional material oriented around ESP-IDF v5.x FreeRTOS APIs and idioms — not a verbatim dump of vendor documentation.

Dataset summary

Item	Value
Format	Hub `Dataset` / JSONL with ChatML `messages` (`system` / `user` / `assistant`)
Total examples	1160
Train	1101
Validation	59 (~5%, stratified by category)
License	MIT
Language	English
Focus	FreeRTOS on ESP32 (ESP-IDF), embedded C, real-time patterns

Each row has a messages list (ChatML) and a meta object (id, category, tags, difficulty, style, APIs, etc.).

Schema

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "meta": {
    "id": "tasks_basic_0042",
    "category": "tasks_basic",
    "tags": ["xTaskCreate", "stack"],
    "difficulty": "intermediate",
    "style": "code_lab",
    "apis": ["xTaskCreate", "vTaskDelay"],
    "esp_idf_version_note": "v5.x conceptual",
    "source": "generated",
    "multi_turn": false
  }
}

Every row has system plus at least one user / assistant pair.
~12% multi-turn (debug-style follow-ups).
Categories cover tasks, SMP/pin-to-core, queues, semaphores/mutexes, event groups, FreeRTOS & ESP timers, stream/message buffers, notifications, ISR/critical sections, heap/heap_caps, ring buffers, hooks/TLS/WDT, design patterns, debug/troubleshoot, and compare/choose.

Load

from datasets import load_dataset

ds = load_dataset("rafuke/freertos_esp32")
print(ds)
print(len(ds["train"]), len(ds["validation"]))
print(ds["train"][0]["messages"][0]["role"])

Fine-tuning sketch (Unsloth + Qwen)

Map messages through the tokenizer chat template:

from datasets import load_dataset

dataset = load_dataset("rafuke/freertos_esp32")

def to_text(batch, tokenizer):
    texts = [
        tokenizer.apply_chat_template(
            msgs,
            tokenize=False,
            add_generation_prompt=False,
        )
        for msgs in batch["messages"]
    ]
    return {"text": texts}

# train_ds = dataset["train"].map(lambda b: to_text(b, tokenizer), batched=True, ...)

Intended use

Supervised fine-tuning (SFT) of instruct models for ESP32 / FreeRTOS Q&A and code assistance.
Evaluation of embedded RTOS knowledge in chat format.

Out of scope / known limitations

Not a substitute for official docs. Prefer Espressif ESP-IDF and FreeRTOS reference manuals for authoritative API contracts, errata, and chip-specific behavior.
Answers target ESP-IDF v5.x conceptual usage; APIs and defaults can differ across IDF versions and SoCs (ESP32, S2, S3, C3, C6, H2, etc.).
Vanilla FreeRTOS vs ESP-IDF FreeRTOS differences (SMP, pin-to-core, TWDT, heap_caps, ring buffers) are called out in spirit, but edge cases may be incomplete.
Generated instructional content may contain occasional inaccuracies; validate critical code on hardware.
Not a full project corpus (no complete sdkconfig / multi-file apps per row).
English only.

License

MIT — see LICENSE in this repository.

Citation

If you use this dataset, please cite the Hub repo:

@misc{freertos_esp32,
  title  = {FreeRTOS ESP32},
  author = {rafuke},
  year   = {2026},
  url    = {https://huggingface.co/datasets/rafuke/freertos_esp32}
}