aab20abdullah's picture
LearnPath GGUF, Modelfile and model card
a58c3b2 verified
|
Raw History Blame Contribute Delete
5.15 kB
---
license: apache-2.0
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
base_model: google/gemma-4-E4B-it
base_model_relation: finetune
library_name: gguf
pipeline_tag: text-generation
language:
- en
- ar
- tr
tags:
- gguf
- ollama
- gemma4
- unsloth
- qlora
---
# LearnPath · Gemma 4 E4B (Q4_K_M GGUF)
Gemma 4 E4B-it fine-tuned for **LearnPath (مسار)**, a study-plan generator that runs entirely on your own
machine: give it a topic, hours per day and a number of days, and it builds a roadmap, a daily to-do list
and free resources for every task.
> نموذج مسار: يحوّل الموضوع والوقت المتاح إلى خطة تعلّم يومية بمصادر مجانية، ويعمل محلياً بالكامل.
>
> Masar modeli: konuyu ve ayırabildiğin zamanı ücretsiz kaynaklarla günlük bir öğrenme planına dönüştürür; tamamen yerel çalışır.
The model does the three language steps of that pipeline and answers with JSON only, in English, Arabic or
Turkish (the learner's language is passed as `en`, `ar` or `tr`).
Scheduling, web search and link checks are ordinary code in the app.
| Task | Gets | Returns |
|---|---|---|
| intake | topic, level, hours per day, days, goal | `topic`, `level`, `total_hours`, `realistic`, `warnings` (in the learner's language) |
| curriculum | topic, level, learning-hour budget | `skills`: ordered `id`, `name`, `hours`, `depends_on`, `search_query` |
| rank | a skill and numbered search results | `indexes` of the best results (never writes links itself) |
## Run with Ollama
```bash
hf download aab20abdullah/learnpath-gemma4-e4b-gguf --local-dir models/
ollama create learnpath -f models/Modelfile
ollama run learnpath
```
`Modelfile` holds the Gemma 4 chat template (thinking off, temperature 0.3, 8k context). The Q4_K_M file is
5.0 GB and runs on CPU or GPU. For dependable JSON, pass the answer's
JSON schema as Ollama's `format`, as the app does.
## Prompt format
The model was trained on these exact prompts; other wording works less reliably. The system prompt is fixed
per task and the user message is filled in:
**intake**
```text
[system]
You are a learning planner. Assess whether the learner's time budget is realistic for the topic and level. Reply with JSON only.
[user]
Topic: Python programming
Level: beginner
Hours per day: 2
Days: 30
Goal: general mastery
total_hours must equal 60. Add warnings (in the learner's language: en) if the budget is too small or too large.
```
**curriculum**
```text
[system]
You are a curriculum designer. Break the topic into ordered skills. Each skill has a short snake_case id, a name, estimated hours, depends_on (ids of earlier skills only) and a search_query for finding free tutorials. Never output URLs. Reply with JSON only, exactly in this shape: {"skills": [{"id": str, "name": str, "hours": number, "depends_on": [str], "search_query": str}]}
[user]
Topic: Python programming
Level: beginner
Total learning hours available: 48.0. Skill hours must sum to about this.
Skill names in language: en. search_query in English.
```
**rank**
```text
[system]
Pick the best free learning resources for a skill from the numbered candidates. Only choose indexes from the list; never invent links. Reply with JSON only.
[user]
Skill: Python loops
Pick up to 2:
0. [video/youtube] Python loops - full tutorial for beginners
1. [article/brave] Best pizza recipes 2026
2. [docs/brave] The for statement - Python documentation
```
## Evaluation
195 held-out examples: 75 topics never seen in training (25 each in English, Arabic and Turkish) for intake and
curriculum, plus 45 rank lists that hide 2 good resources among 4 distractors. Greedy decoding, no schema constraint, the same
automatic checks for both models. An example passes only if every check holds: valid JSON and schema, total
hours, language of names and warnings, dependency order, no URLs, skill count, and picking only good resources.
Measured on the 4-bit LoRA model before the GGUF export.
| Task | Examples | JSON valid: base → tuned | Pass: base → tuned |
|---|---|---|---|
| intake | 75 | 100% → 100% | 0% → **100%** |
| curriculum | 75 | 100% → 100% | 87% → **95%** |
| rank | 45 | 100% → 100% | 0% → **100%** |
| **all** | 195 | | 33% → **98.0%** |
The base model's 0% on intake and rank comes mostly from wrong JSON key names. Passing the schema as Ollama's
`format` fixes the keys, so in the app the gain is mainly in the content checks.
## Training
- QLoRA (4-bit base) with Unsloth on one Kaggle T4: LoRA r=16, alpha=32, language
layers only, loss on the answers only.
- Data: 5,701 synthetic examples (curriculum 2,700, rank 1,800, intake 1,201; English 1,911, Turkish 1,901, Arabic 1,889) written by Gemma 4 26B-A4B running locally with llama.cpp, kept only when they pass the same automatic checks. The eval topics are left out.
- 708 steps in 169 min, final train loss 0.198.
## Limitations
- Made for these three prompts; it is not a general chat model.
- Skill hours can drift from the budget; the app rescales them to fit.
- It can only rank the search results it is given.