Antlia-Learn-7B

Summary

Antlia-Learn-7B is a 7B-parameter instruction-tuned model for instructional design. It drafts concise short micro-courses and complex long-form course outlines, including learning outcomes, section structure, short practice activities, and lightweight checkpoints. This model is released under the GNU Affero General Public License v3.0 (AGPL-3.0).

Intended use

• Draft micro-courses with one goal, a 3–4 point outline (minutes in parentheses), and 1–2 short practice tasks.
• Draft long-form course outlines with section objectives, activities, checkpoints, and assessment ideas.
• Rewrite learning objectives in plain language and propose short Socratic prompts.

Out of scope Safety-critical or regulated domains (medical, legal, financial), processing of personal/regulated data, and unsupervised deployment without human review.

Inputs and outputs

Input: English prompts describing topic and scope (e.g., “Create a micro-course on threading a needle” or “Design a 3–4 hour course on building a car engine for advanced beginners”). Output: English text: for micro-courses, a goal + outline + practice tasks; for longer courses, multi-section plans with objectives, activities, checkpoints, and assessment ideas.

Recommended decoding

The repository includes generation_config.json tuned for concise, low-repetition drafting:

• do_sample: true
• temperature: 0.35
• top_p: 0.85
• top_k: 40
• repetition_penalty: 1.3
• no_repeat_ngram_size: 5
• max_new_tokens: 120
• min_new_tokens: 16

Model details

• Architecture: Decoder-only transformer, ~7B parameters.
• Precision: Standard HF weights; quantized inference supported downstream.
• Alignment: Supervised instruction-tuning focused on clear, compact educational scaffolds; no RLHF in this release.

Training

Supervised instruction-tuning across staged passes to encourage concise goals, minute-budgeted outlines, checkpoints, and practice prompts. Exact mixtures, volumes, and hyperparameters are proprietary.

Limitations and risks

• Not a knowledge base; facts may be incomplete or outdated.
• Tends toward brevity; for richer syllabi, include explicit section counts, activities, and assessment style.
• Can produce generic phrasing for very broad prompts—be specific.

Safety

All outputs require human review for accuracy and appropriateness. Do not use to generate harmful, discriminatory, or unsafe content.

License

• Weights: GNU Affero General Public License v3.0
• Use: Your use must comply with the license and any applicable upstream licenses.

Third-party notice (required)

This model’s training process made use of content from BAAI/Infinity-Instruct, licensed under Creative Commons Attribution-ShareAlike 4.0 International. License: https://creativecommons.org/licenses/by-sa/4.0/ Dataset page: https://huggingface.co/datasets/BAAI/Infinity-Instruct

Contact

Maintainer: OOMU — antlia@oomu.ai