Upload README.md with huggingface_hub

README.md

@@ -1,35 +1,19 @@
 ---
 {}
 ---
----
-            license: apache-2.0
-            library_name: transformers
-            base_model: google/gemma-2b
-            tags:
-              - text-generation
-              - fine-tuned
-              - pdf-grounded
-              - zero-hallucination
-              - domain-specific
-            language:
-              - en
-            pipeline_tag: text-generation
-            ---
 
-            # 📂 Solvrays Llm (Precision Grounded)
+            # 📂 Solvrays Llm (Ground-Truth Precise)
 
             ## 🌟 Overview
-            This model is a high-precision version of **Gemma 2B**, fine-tuned for **strict document retrieval and analysis**. It has been conditioned through deterministic instruction-grounding templates to minimize hallucinations and prioritize facts extracted directly from provided source materials.
-
-            ### 🛠 Core Advancements
-            - **Universal Document Grounding**: Optimized for both technical and non-technical corpora.
-            - **Negative Constraint Training**: Trained to prioritize "Not Documented" over guessing when information is missing.
-            - **Deterministic Inference**: Configured for greedy decoding to ensure factual consistency.
-            - **High Continuity**: Trained with 128-token chunk overlap to preserve context across page boundaries.
+            This is a specialized fine-tuned version of **Gemma 2B**, optimized for **High-Precision Document Retrieval**. It has been trained using strict grounding templates to ensure zero-hallucination and deterministic factual responses.
 
-            ## 💻 Precision Quick Start
-            To ensure zero-hallucination, use the following **Universal Grounding Prompt**:
+            ## 🛠 Key Advanced Features
+            - **Zero-Hallucination Mode**: Deterministic greedy decoding by default.
+            - **Negative Constraint Awareness**: Trained to avoid guessing when information is missing.
+            - **Domain Agnostic**: Works for any technical or non-technical PDF provided as context.
+            - **Standalone Conversion**: Fully merged FP16 weights for production deployment.
 
+            ## 💻 Quick Start (Inference)
             ```python
             from transformers import AutoTokenizer, AutoModelForCausalLM
             import torch
@@ -38,37 +22,22 @@
             tokenizer = AutoTokenizer.from_pretrained(model_id)
             model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.float16)
 
-            # The EXACT prompt used during fine-tuning grounding
             instruction = "Analyze the following document and provide a precise, factual response based strictly on the content provided. If the information is not present, you must state that it is not documented."
-            document_source = "Your_Document_Name.pdf"
-            query = "Explain the key requirements from the document."
-
             prompt = f"### Instruction: {instruction}
-### Source: {document_source}
-### Content: {query}
+### Source: Document_Name.pdf
+### Content: Your Query Here
 ### Verified Response:"
-            inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-
-            with torch.no_grad():
-                outputs = model.generate(
-                    **inputs, 
-                    max_new_tokens=256, 
-                    do_sample=False,       # Absolute precision mode
-                    repetition_penalty=1.5
-                )
 
+            inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+            outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False, repetition_penalty=1.5)
             print(tokenizer.decode(outputs[0], skip_special_tokens=True).split("### Verified Response:")[-1].strip())
             ```
 
-            ## 📊 Training Methodology
-            - **Model Foundation**: google/gemma-2b
-            - **Technique**: QLoRA with 4-bit NormalFloat (nf4) quantization.
-            - **Rank (r)**: 16 | **Alpha**: 32 (Balanced capacity/stability).
-            - **Epochs**: 5 (Intensive fact-reinforcement).
-            - **Hardware**: Optimized for NVIDIA L4/V100/A100 environments.
-
-            ## 📜 License & Usage
-            Usage is governed by the Apache-2.0 license and the Gemma Prohibited Use Policy. Ideal for internal technical summary, compliance checking, and document search tasks.
+            ## 📊 Training methodology
+            - **Base Model**: google/gemma-2b
+            - **Quantization**: 4-bit (NormalFloat4)
+            - **LoRA Config**: r=16, alpha=32, target_modules=All linears
+            - **Epochs**: 5 (Intensive Reinforcement)
 
             --- 
-            **Professionally Fine-tuned by Bibek Lama Singtan**
+            **Fine-tuned by Bibek Lama Singtan**