| --- |
| license: apache-2.0 |
| pretty_name: Epoch Eli Dataset v3 — S-Tier Coding & Agentic Fine-Tuning Corpus |
| language: |
| - en |
| task_categories: |
| - text-generation |
| tags: |
| - code |
| - agentic-coding |
| - sharegpt |
| - alpaca |
| - vision-language |
| - qwen |
| - epoch |
| --- |
| |
| # Epoch Model Suite 1 — Eli Dataset v3 |
|
|
| Dataset curated for fine-tuning **Eli** (~3-7B dense coding model, based on Qwen 3-4B). |
|
|
| ## Dataset Composition |
|
|
| - **Total S-Tier Samples**: 23,138 pairs |
| - **Backend Code (40.5%)**: 9,360 clean pairs across Linux, Kubernetes, Tokio, FastAPI, Hyper, Gin. |
| - **Frontend Code & Design Systems (22.9%)**: 5,290 pairs across shadcn/ui, Radix, Mantine, Tailwind, Stripe, Vercel, Apple HIG specs. |
| - **Eli Personality (30.2%)**: 6,988 pairs (1,747 unique Q&As oversampled 4x) capturing Eli's casual, direct, fluff-free persona. |
| - **Fable-5 Pi Agent Traces (4.3%)**: 1,000 real agentic execution traces (CoT reasoning, Bash/Read/Edit actions). |
| - **Agentic Open-Ended Apps (2.2%)**: 500 unique single-file apps and microservices. |
| - **Eli-VL Multimodal**: 192 UI WebP screenshot-to-code pairs (`training-data-eli-vl.jsonl`). |
|
|
| ## Files Included |
|
|
| - `training-data.jsonl`: Alpaca format (`instruction`, `output`). |
| - `training-data-sharegpt.jsonl`: ShareGPT / ChatML format (`conversations`). |
| - `training-data-eli-vl.jsonl`: Multimodal screenshot-to-code pairs. |
| - `dataset-stats.json`: Automated statistical breakdown. |
|
|
