CoolFace
Datasetpublic

grandekos/AIIA_DATASET

OurAI Training Datasets Fine-tuning datasets for meta-llama/Llama-3.2-3B-Instruct (QLoRA) using the OurAI framework tool-calling patterns. Files File Examples Size Description ourai_all_data.jsonl 2,355 37 MB Unified dataset from all .dbk session files (OurAI/history/ + playground) + t20 HISTORY.md. Developer/coding sessions only, deduplicated by content hash. ourai_new_data.jsonl 148 3.2 MB Legacy: parsed from HISTORY.md transcripts (t7–t17, KosDB).… See the full description on the dataset page: https://huggingface.co/datasets/grandekos/AIIA_DATASET.

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes12downloads
Dataset Card

OurAI Training Datasets

Fine-tuning datasets for meta-llama/Llama-3.2-3B-Instruct (QLoRA) using the OurAI framework tool-calling patterns.

Files

FileExamplesSizeDescription
ourai_all_data.jsonl2,35537 MBUnified dataset from all .dbk session files (OurAI/history/ + playground) + t20 HISTORY.md. Developer/coding sessions only, deduplicated by content hash.
ourai_new_data.jsonl1483.2 MBLegacy: parsed from HISTORY.md transcripts (t7–t17, KosDB). Used for v8 training.
ourai_clean.jsonl3152.6 MBLegacy: filtered/cleaned combination of earlier data. Used for v7 training.
ourai_train.jsonl3151.7 MBAlias for ourai_clean.jsonl.
ourai_merged.jsonl—1.9 MBLegacy: merged combination of earlier datasets.

Format (JSONL)

Each line is a Llama 3.2 Instruct-format conversation:

json
{"text": "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system}<|eot_id|>...<|eot_id|>"}

Fields: {"text": "<full_chat_template>"} — compatible with train_ourai.py.

Tool calls and results are embedded inline in the assistant content (OurAI framework convention).

Sources

.dbk session files (JSONL format)

  • —OurAI/history/ — 651 files, 714 MB (primary source, 312 developer sessions parsed)
  • —contabo1_aiia_history/ — 10 session files (sessions 315–340)
  • —kosgen1_aiia_history/ — 26 session files (sessions 314–341)
  • —contabo1_aiia_host76_history/ — 15 session files (sessions 314–323)

Each .dbk file is a New Line Delimited JSON recording of a real agent interaction session with fields: role, content, sessionId, rowId, timestamp. Only developer/coding sessions are included (filtered by system prompt keywords).

HISTORY.md files

  • —collecting_program_data/t20/ — parsed from HTML-comment JSON blobs
  • —(Earlier sessions t7–t17, KosDB are in ourai_new_data.jsonl)

Regeneration

bash
python ../../prepare_all_ourai_data.py

Usage

bash
# Train on the unified dataset
cd data/processed_datasets/ourai
python ../../../OurAI/OurAI/examples/train_ourai.py \
  --data ourai_all_data.jsonl \
  --output-dir models/ourai_v9 \
  --epochs 3 --lr 2e-4 \
  --lora-r 16 --lora-alpha 32 \
  --max-seq-length 2048