Samrish2009/SAM-AI-Reasoning-v4

SAM-AI Reasoning v4 (Parallax)

SAM-AI Reasoning v4 is a 14B parameter reasoning adapter developed by Parallax (Samrish). It is fine-tuned using Group Relative Policy Optimization (GRPO) with rule-based verifiable rewards on top of deepseek-ai/DeepSeek-R1-Distill-Qwen-14B.

Model Overview

• Base Architecture: DeepSeek-R1-Distill-Qwen-14B (Qwen2.5 14B transformer backbone)
• Adapter Type: LoRA (Rank = 16, Alpha = 32, Dropout = 0.05)
• Target Modules: All linear projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj)
• Training Method: GRPO (Group Relative Policy Optimization) with format verification, multi-step math/code reward checks, and reasoning traces (<think>...</think>).
• Trained Parameters: 137.7 MB adapter safetensors.

Usage with PEFT & Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B"
adapter_name = "Samrish2009/SAM-AI-Reasoning-v4"

tokenizer = AutoTokenizer.from_pretrained(base_model_name)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, adapter_name)

prompt = "<｜User｜>Solve step by step: Prove that for any positive integer n, n^3 + 2n is divisible by 3.<｜Assistant｜><think>\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.6, top_p=0.95)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Curriculum & Training Objectives

SAM-AI v4 incorporates:

1. Verifiable Reasoning: Strict formatting enforcement and multi-step deduction traces.
2. Abstract Spatial Logic: Curriculum drawn from ARC inductive reasoning patterns.
3. Mathematical Derivations: Step-by-step rigorous proof generation.
4. Code Execution & Verification: Synthesizing verifiable Python programs.

Developed by

• Team: Parallax
• Lead Developer: Samrish