bf369 commited on
Commit
92526ef
·
verified ·
1 Parent(s): 7a02192

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +60 -15
README.md CHANGED
@@ -1,24 +1,69 @@
1
- ---
 
 
 
 
2
  tags:
 
 
 
 
 
3
  - gguf
4
- - llama.cpp
5
- - unsloth
6
-
 
 
 
 
7
  ---
8
 
9
- # BeejX-Gemma2-2B-Hindi-SFT - GGUF
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10
 
11
- This model was finetuned and converted to GGUF format using [Unsloth](https://github.com/unslothai/unsloth).
 
12
 
13
- **Example usage**:
14
- - For text only LLMs: **llama-cli** **--hf** repo_id/model_name **-p** "why is the sky blue?"
15
- - For multimodal models: **llama-mtmd-cli** **-m** model_name.gguf **--mmproj** mmproj_file.gguf
16
 
17
- ## Available Model files:
18
- - `gemma-2-2b-it.Q4_K_M.gguf`
19
 
20
- ## Ollama
21
- An Ollama Modelfile is included for easy deployment.
22
 
23
- ## Note
24
- The model's BOS token behavior was adjusted for GGUF compatibility.
 
 
1
+ """---
2
+ language:
3
+ - hi
4
+ - en
5
+ - hinglish
6
  tags:
7
+ - agriculture
8
+ - farming
9
+ - india
10
+ - kisan
11
+ - gemma
12
  - gguf
13
+ - text-generation-inference
14
+ - on-device
15
+ - android
16
+ base_model: unsloth/gemma-2-2b-it-bnb-4bit
17
+ license: gemma
18
+ pipeline_tag: text-generation
19
+ quantized_by: bf369
20
  ---
21
 
22
+ # BeejX-Gemma2-2B-Hindi-SFT (GGUF)
23
+
24
+ **BeejX** is a lightweight, offline-capable AI assistant specifically fine-tuned for **Indian Agriculture** (Bhartiya Krishi).
25
+ It is designed to run on low-end Android smartphones (4GB+ RAM) using the GGUF format, making it accessible to farmers without active internet connections.
26
+
27
+ ## Model Details
28
+ - **Base Model:** [Gemma 2 2B Instruct](https://huggingface.co/google/gemma-2-2b-it) (Google)
29
+ - **Fine-tuning:** Supervised Fine-Tuning (SFT) using [Unsloth](https://github.com/unslothai/unsloth) on 1,000+ samples of Indian agriculture Q&A.
30
+ - **Language:** Hindi (Devanagari), Hinglish (Roman Hindi), and English.
31
+ - **Format:** GGUF (Quantized to `Q4_K_M` for best Mobile balance).
32
+ - **File Size:** ~1.5 GB (Fits easily on modern Android phones).
33
+
34
+ ## Capabilities
35
+ - **Crop Advice:** Detailed guides for Wheat (Gehu), Rice (Dhaan), Sugarcane (Ganna), Maize (Makka), etc.
36
+ - **Disease Identification:** Describes symptoms and organic/chemical remedies (e.g., Yellow Rust, Termites, Blight).
37
+ - **Localized Knowledge:** Understands Indian seasons (Rabi, Kharif, Zaid) and local fertilizers (DAP, Urea, NPK).
38
+ - **Safety:** Trained to refuse non-agriculture questions to keep the app focused.
39
+
40
+ ## How to Run on Android (Flutter)
41
+ This model is optimized for the `llama.cpp` ecosystem.
42
+
43
+ 1. **Download:** Get the `unsloth.Q4_K_M.gguf` file from the [Files tab](./tree/main).
44
+ 2. **Move to Device:** Place the file in your App's `ApplicationDocumentsDirectory`.
45
+ 3. **Flutter Integration:** Use a library like `llama_cpp_dart` or `fllama`.
46
+ 4. **Recommended Parameters:**
47
+ - `context_size`: 2048
48
+ - `temperature`: 0.7
49
+ - `system_prompt`:
50
+ ```text
51
+ You are BeejX, an expert Indian agriculture assistant.
52
+ Answer in the same language as the user (Hindi/Hinglish).
53
+ Keep answers short, practical, and easy for a farmer to understand.
54
+ ```
55
 
56
+ ## How to Run on PC (Ollama)
57
+ You can test this model on your laptop using Ollama before deploying to mobile.
58
 
59
+ ## Limitations & Disclaimer
60
+ AI Advice: This is an AI model. While trained on agricultural data, it can hallucinate. Users should verify chemical dosages with a local Krishi Vigyan Kendra (KVK) expert.
 
61
 
62
+ Language: Performance is best in Hindi and Hinglish. Pure regional dialects (like Punjabi/Bhojpuri) may have lower accuracy.
 
63
 
64
+ ## Credits
65
+ Finetuned by **Bhashkar Fulara** using Unsloth .
66
 
67
+ ```bash
68
+ # Run directly from Hugging Face
69
+ ollama run hf.co/bf369/BeejX-Gemma2-2B-Hindi-SFT