Upload 6 files


LICENSE

@@ -0,0 +1,6 @@
+@misc{abirhinv1,
+  author       = {Abir Maheshwari},
+  title        = {ABIRHINv1: Pure Hindi Model from Scratch},
+  year         = {2026},
+  url          = {https://huggingface.co/AbirMaheshwari/ABIRHINv1}
+}

README.MD

@@ -0,0 +1,138 @@
+\# ABIRHINv1
+
+
+
+\*\*Pure Hindi Language Model – Built Entirely From Scratch\*\*
+
+
+
+\*\*Version 1\*\* – Created by Abir Maheshwari in February 2026 using only Google Colab free tier (T4 GPU). ≈100 million parameters.
+
+
+
+\### About the Model
+
+ABIRHINv1 is the second member of the ABIR Indic SLM Family (after Marathi).  
+
+It is a \*\*decoder-only causal language model\*\* built \*\*100% from scratch\*\*:
+
+
+
+\- Random weight initialization (no pretrained checkpoints or base models)
+
+\- Custom architecture using PyTorch `nn.TransformerDecoder` layers
+
+\- Custom tokenizer trained only on Hindi data (Byte-level BPE from zero – no inheritance from any existing tokenizer)
+
+\- Trained exclusively on Hindi text (IndicCorpV2), Romanized Hindi (Bhasha-Abhijnaanam), creator personality, and translation pairs
+
+
+
+Generates fluent Hindi (Devanagari), understands Romanized input, basic English → Hindi translation.
+
+
+
+\### Purpose \& Motive
+
+\*\*Purpose\*\*: Tiny, offline Hindi AI for millions in North India – no internet, low-end devices.  
+
+\*\*Motive\*\*: Show anyone can build Indic models from scratch. Empowering Hindi speakers in the AI era.
+
+
+
+\### Target Audience
+
+\- Hindi families \& kids
+
+\- North India users (daily chat, news, forms)
+
+\- Students, writers, teachers
+
+\- Offline developers
+
+
+
+\### Capabilities (Version 1)
+
+\- Fluent Hindi generation
+
+\- Romanized understanding ("Main kya karun?")
+
+\- English → Hindi translation
+
+\- Knows creator: \*\*Abir Maheshwari from Mumbai\*\*
+
+\- ~400–500 MB size – offline fast
+
+
+
+\### Use Cases
+
+1\. Family Hindi chatbot
+
+2\. Stories \& poems
+
+3\. Writing help
+
+4\. Quick translation
+
+5\. Offline learning
+
+
+
+\### Creator Information
+
+\*\*Created by\*\*: Abir Maheshwari (Mumbai, Maharashtra, India)  
+
+\*\*Writer • Programmer • Entrepreneur • Artist\*\*
+
+
+
+\*\*Follow me:\*\*  
+
+\- X / Twitter: \[@AbirMaheshwari](https://x.com/AbirMaheshwari)  
+
+\- Instagram: \[@anantraga31](https://instagram.com/anantraga31)  
+
+\- LinkedIn: \[Abir Maheshwari](https://linkedin.com/in/abirmaheshwari)
+
+
+
+\*\*Model says\*\*: "मैं ABIRHINv1 हूँ। मेरे निर्माता अभीर महेश्वरी हैं।"
+
+
+
+\### Technical Details
+
+\- Architecture: Custom decoder-only (10 layers, 640 dim, 10 heads, GELU, learnable pos)
+
+\- Parameters: ≈100 million
+
+\- From scratch: Yes
+
+\- Tokenizer: Byte-level BPE from zero (32k vocab)
+
+\- Dataset: IndicCorpV2 (hin\_Deva), Bhasha-Abhijnaanam Hindi, custom pairs
+
+\- Compute: Colab free T4 (~1–2 hours)
+
+
+
+\### Limitations
+
+Small model – basic fluency, short context, no real-time knowledge.
+
+
+
+\### How to Use
+
+```python
+
+from transformers import pipeline
+
+
+
+pipe = pipeline("text-generation", model="AbirMaheshwari/ABIRHINv1")
+
+print(pipe("मेरे निर्माता कौन हैं?", max\_new\_tokens=80)\[0]\["generated\_text"])
+

config.json

@@ -0,0 +1,16 @@
+{
+  "architectures": [
+    "ABIRForCausalLM"
+  ],
+  "dropout": 0.1,
+  "dtype": "float32",
+  "hidden_size": 640,
+  "intermediate_size": 2560,
+  "max_position_embeddings": 512,
+  "model_type": "abir-slm",
+  "num_heads": 10,
+  "num_layers": 10,
+  "transformers_version": "5.0.0",
+  "use_cache": false,
+  "vocab_size": 32000
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db193b7f8dd126b7345e46c25df6d6a1f26c3c5c205fd57721b75b4a9190105f
+size 427800504

tokenizer_config.json

@@ -0,0 +1,9 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<bos>",
+  "eos_token": "<eos>",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<pad>",
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": "<unk>"
+}