SanatanSinghVishen/sift-1b-gguf

Sift-1B: Deterministic Function Calling & Intent Routing SLM

Sift-1B is a hyper-specialized 1.5B-parameter Small Language Model (SLM) engineered specifically to sift structured JSON parameters out of unstructured user prompts and shift routing workloads away from expensive cloud LLMs.

Trained via a two-stage Supervised Fine-Tuning (SFT) + Direct Preference Optimization (DPO) pipeline on top of Qwen/Qwen2.5-1.5B-Instruct, Sift-1B strips away all conversational fluff ("Sure! Here is your JSON:") and outputs strict, machine-readable JSON on the very first attempt.

• —Developed & Designed by: Sanatan Singh
• —Base Architecture: `Qwen/Qwen2.5-1.5B-Instruct`
• —Model Type: Causal Language Model + QLoRA Adapter + GGUF Quantization (q4_k_m)
• —License: MIT
• —GitHub Repository: SanatanSinghVishen/Sift-1B
• —GGUF Quantized: `SanatanSinghVishen/sift-1b-gguf`
• —DPO Golden Adapter: `SanatanSinghVishen/sift-1b-dpo`
• —SFT Adapter: `SanatanSinghVishen/sift-1b-sft`

📊 Comprehensive Benchmark Results

Evaluated across 50 holdout test cases containing UNSEEN function schemas (not present in training) comparing Base Qwen vs. SFT vs. DPO (Golden Release Checkpoint-750):

⚙️ Technical Specifications & GGUF Metadata

GGUF Architecture Parameters

• —Quantization Format: q4_k_m (4-bit medium K-quantization)
• —GGUF File Size: 940.4 MB (0.94 GB)
• —Context Length (`n_ctx`): 32,768 tokens
• —Embedding Dimensions (`n_embd`): 1,536
• —Feed-Forward Network Dimension (`n_ff`): 8,960
• —Attention Heads (`n_head`): 12
• —Key-Value Heads (`n_head_kv`): 2 (Grouped-Query Attention / GQA)
• —RMS Norm Epsilon: 1e-6
• —RoPE Theta Base (`rope_theta`): 1,000,000.0
• —Vocabulary Size: 151,936 tokens (ChatML format)

Training Hyperparameters

Stage 1: Supervised Fine-Tuning (SFT)

• —Dataset: 10,000 ChatML multi-turn function calling samples derived from Salesforce/xlam-60k
• —Quantization: QLoRA 4-bit NF4 (load_in_4bit=True)
• —LoRA Rank ($r$): 16 | LoRA Alpha ($\alpha$): 32 | Dropout: 0
• —Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• —Batch Size: 1 per device | Gradient Accumulation: 8 (Effective batch size = 8)
• —Learning Rate: 2.0e-4 (Cosine schedule, warmup_ratio=0.05)
• —Epochs: 2 (2,500 total steps) | Optimizer: adamw_8bit

Stage 2: Direct Preference Optimization (DPO)

• —Dataset: Synthetically mutated preference pairs ($yw$ = clean JSON, $yl$ = markdown fluff/hallucinations)
• —DPO Temperature ($\beta$): 0.1 | Loss Type: sigmoid
• —Learning Rate: 5.0e-6 (Cosine schedule, warmup_ratio=0.1)
• —Epochs: 1 (2,500 total steps, golden checkpoint at step 750)

🚀 Quick Start Code Examples

1. Ollama (Local CLI & Server)

# Run directly from Hugging Face Hub:
ollama run hf.co/SanatanSinghVishen/sift-1b-gguf