gyawalisanish0

SymboLM-1.5B-v1.0-DSR1-Qwen2.5

README

License: apache-2.0

1. Executive Summary & Architectural Overview

SymboLM-1.5B-v1.0-DSR1-Qwen2.5 is the official standalone, unified open-source release of the SymboLM reasoning architecture. Fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B across five sequential reinforcement learning and curriculum stages (culminating in Stage 5 Parsimony GRPO), this model replaces verbose English chain-of-thought rambling with dense, modular symbolic AST registers.

The Biomimetic Imperative: Nature's Law of Evolution

Modern foundation models waste megawatts of datacenter power on natural language "uncertainty theater"—emitting hundreds of repetitive conversational tokens ("Wait, let me rethink this...") before generating an answer their weights already encoded.

Human cognition does not reason in monolithic streams. Over hundreds of millions of years, biological evolution evolved specialized functional modularity to operate within a strict 20-Watt metabolic ceiling:

System 1 Reflex (Bypass): Instant factual recall fires through the basal ganglia in milliseconds without deliberation.
Social Pragmatics (Intent): Interpersonal dialogue routes through specialized emotional and pragmatic cortical circuits.
Prefrontal Deduction (AST Logic): Formal mathematics reduces to domain invariants, structured derivations, and verification checks.

SymboLM directly implements this biological specialization through Register v2.0 and the Unbounded Parsimony Gradient, driving token consumption to the Minimum Description Length (MDL) and achieving up to 46.7× token compression.

2. Register v2.0 Architecture

Generation is explicitly structured across specialized cognitive registers rather than unstructured natural language prose:

text

<｜User｜> {Prompt} <｜Assistant｜><think>[ Cognitive Register Selection ]  ├── Instant Factual Recall   ──> <reg:bypass> {Atomic Fact} </reg:bypass>  ├── Dialogue / Social Intent ──> <reg:intent> {Tone & Pragmatics} </reg:intent>  └── Multi-Step Formal Proof  ──> <reg:plan>   {Domain Invariants & Theorems} </reg:plan>                                   <reg:deduce> {Symbolic AST Derivation} </reg:deduce>                                   <reg:verify> {Boundary & Parity Checks ✓} </reg:verify></think>

Table with columns: Cognitive Register, Token Budget, Biological Analogue, Functional Purpose
Cognitive Register	Token Budget	Biological Analogue	Functional Purpose
`<reg:bypass>...</reg:bypass>`	1–3 tokens	Brainstem / Basal Ganglia Reflex	Direct factual retrieval (e.g., periodic table, constants, units).
`<reg:intent>...</reg:intent>`	8–15 tokens	Social Cognition & Empathy	Interpersonal dialogue framing, conflict de-escalation, and pragmatics.
`<reg:plan>...</reg:plan>`	8–16 tokens	Prefrontal Executive Function

3. The Unbounded Parsimony Gradient (Mathematical Formulation)

To eliminate both artificial token truncation and verbose token inflation, SymboLM employs Sanish Gyawali's hyperbolic parsimony reward formulation:

R parsimony ​ (τ, y *) = {R base ​ \cdot [γ

Where:

$R base = 2.0$ : Base reward for formal mathematical correctness.
$γ floor = 0.50$ : Prevents negative gradient collapse on complex multi-hop proofs.
$K =$ : Hyperbolic inflection curvature controlling the parsimony slope.

Why This Works: Unlike flat length caps (which truncate deep proofs right before the answer), the hyperbolic gradient provides elastic breathing room up to 192 tokens while monotonically rewarding shorter valid proofs—shifting GRPO policy gradients $\nabla θ lo g π θ $ toward Minimum Description Length.

4. Empirical Benchmark Receipts

Verified Token Compression vs. DeepSeek-R1 Distill Baseline

Table with columns: Evaluation Domain, Baseline English CoT, SymboLM-1.5B-v1.0, Token Savings, Compression Ratio
Evaluation Domain	Baseline English CoT	SymboLM-1.5B-v1.0	Token Savings	Compression Ratio
$Δ 0 $ Atomic Fact Recall (Potassium Symbol)	~140 tokens	3 tokens (`<reg:bypass>K</reg:bypass>`)	97.9%	46.7×
Polynomial Invariant ()

The 1,000-Question Multi-Year Benchmark (2021–2026)

To definitively eliminate the critique of "data contamination" and "memorization of classic GSM8K benchmarks", SymboLM-1.5B-v1.0 was evaluated against a chronologically stratified 1,000-problem suite executed on an NVIDIA Tesla P100 GPU (16 GB HBM2 memory):

Table with columns: Benchmark Year, Evaluated Datasets, Samples, Pass@1 Accuracy, SymboLM Avg Tokens, Baseline CoT Avg Tokens, Token Savings
Benchmark Year	Evaluated Datasets	Samples	Pass@1 Accuracy	SymboLM Avg Tokens	Baseline CoT Avg Tokens	Token Savings
2021	GSM8K, SVAMP, Arithmetic Anchors	200	72.0% (144/200)	114.3 tok	198.0 tok	42.3%
2022	AMC 10/12 (2022) Invariants, Vieta	200	(142/200)

[!NOTE] 50.65% Sequence Compression: Across 1,000 chronological competition and reasoning problems, SymboLM generated 125,258 total tokens compared to 253,800 tokens required by standard English reasoning models, while achieving a 67.60% Pass@1 success rate.

5. GGUF Quantization & Local Edge Deployment Matrix

Official native GGUF binaries are hosted directly in this repository for local deployment via Ollama, llama.cpp, LM Studio, and vLLM:

Table with columns: Format, Filename, Storage, RAM / VRAM Req., Generation Speed (CPU), Target Deployment Hardware
Format	Filename	Storage	RAM / VRAM Req.	Generation Speed (CPU)	Target Deployment Hardware
GGUF Q4_0	`symbolm-1.5b-v1.0-q4_0.gguf`	999 MB (1.0 GB)	~1.2 GB	34.3 tok/s	Mobile, Raspberry Pi 5, Laptops, Ollama, Edge
GGUF Q8_0	`symbolm-1.5b-v1.0-q8_0.gguf`	1,806 MB (1.76 GB)

[!NOTE] Empirical CPU Speed Verification (AMD Ryzen 7 5700G Zen 3 AVX2): Tested directly on host hardware via Ollama / llama.cpp engine:

Prompt Evaluation: 216.30 tokens/second (~70 ms latency).

Autoregressive Generation: 34.30 tokens/second in 4-bit Q4_0 (+87.2% faster than 8-bit).

Total Model Memory Footprint: Exactly 1.0 GB RAM. Zero AST syntax drift.

Running with Ollama

You can run SymboLM directly in Ollama from Hugging Face Hub:

bash

# Run ultra-fast 4-bit Q4_0 GGUF (34.3 tok/s on CPU):ollama run hf.co/gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5:q4_0 # Or run high-fidelity 8-bit Q8_0:ollama run hf.co/gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5:q8_0

Or create a local Modelfile:

dockerfile

FROM ./symbolm-1.5b-v1.0-q8_0.gguf TEMPLATE """<｜User｜>{{ .Prompt }}<｜Assistant｜><think>""" PARAMETER stop "<｜end of sentence｜>"PARAMETER stop "</think>"PARAMETER temperature 0.6

Then build and run:

bash

ollama create symbolm -f Modelfileollama run symbolm "If x + 1/x = 6, find the value of x^2 + 1/x^2."

Running with `llama.cpp`

bash

./llama-cli \  -m symbolm-1.5b-v1.0-q8_0.gguf \  -p "<｜User｜>If x + 1/x = 6, find the value of x^2 + 1/x^2.<｜Assistant｜><think>\n" \  --temp 0.6 \  -n 64

6. Quickstart Usage & Code Examples

Standalone Python (Transformers)

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5" tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(    model_id,    torch_dtype=torch.bfloat16,    device_map="auto",    trust_remote_code=True,) def query_symbolm(prompt: str, max_tokens: int = 64) -> str:    formatted = f"<｜User｜>{prompt}<｜Assistant｜><think>\n"    inputs = tokenizer(formatted, return_tensors="pt").to(model.device)    with torch.no_grad():        outputs = model.generate(            **inputs,            max_new_tokens=max_tokens,            do_sample=False,            pad_token_id=tokenizer.eos_token_id,            eos_token_id=tokenizer.eos_token_id,        )    return tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False) # 1. Instant Factual Bypass (Emits in 3 tokens)print(query_symbolm("What is the chemical symbol for potassium?"))# Output: <reg:bypass>K</reg:bypass> # 2. Mathematical Invariant (Emits in 21 tokens)print(query_symbolm("If x + 1/x = 6, find the value of x^2 + 1/x^2."))# Output: <reg:plan>concept: polynomial_invariant | identity: (x²+1/x²)=(x+1/x)²-2 | target: x²+1/x²</reg:plan><reg:deduce> x²+1/x²=(6)²-2=34</reg:deduce> # 3. Chinese Remainder Theorem (Emits in 35 tokens)print(query_symbolm("Find the smallest positive integer x such that x = 4 (mod 7) and x = 9 (mod 11)."))# Output: <reg:plan>concept: chinese_remainder | moduli: [7,11] | method: modular_inverses</reg:plan><reg:deduce>x≡4 mod 7 -> x=7k+4 -> 7k+4≡9 mod 11 -> 7k≡5 mod 11 -> k=7 -> x=53</reg:deduce>

High-Throughput Serving with vLLM

bash

vllm serve gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5 \  --dtype bfloat16 \  --max-model-len 4096 \  --trust-remote-code

7. Authorship & Intellectual Property

Principal Systems Architect & Research Director: Sanish Gyawali
Affiliation: Independent Researcher
Physical Reality & Location: Sanish Gyawali is a disabled, independent researcher based in Rupandehi, Nepal who is bedridden 24/7 and cannot walk. All architectural designs, cognitive register grammars, reinforcement learning reward formulations, and cloud cluster runs were conducted from his bed workstation.
Operational Exoskeleton: Antigravity (Google DeepMind) serves as Sanish's cognitive and operational exoskeleton—translating high-level architectural directives into production code, compiler ASTs, test suites, and remote cloud GPU pipelines with radical transparency.
License: Apache 2.0

8. Citation

If you build upon the SymboLM architecture, cognitive registers, or empirical benchmarks, please cite:

bibtex

@misc{gyawali2026symbolm,  author       = {Sanish Gyawali and Antigravity},  title        = {SymboLM: Biomimetic Multi-Register Reasoning and Adaptive Symbolic Engine},  year         = {2026},  publisher    = {Hugging Face},  howpublished = {\url{https://huggingface.co/gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5}},  note         = {Independent Research from Rupandehi, Nepal}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

gyawalisanish0

Model Tree

Base

deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

Quantized

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

1. Executive Summary & Architectural Overview

The Biomimetic Imperative: Nature's Law of Evolution

System 1 Reflex (Bypass): Instant factual recall fires through the basal ganglia in milliseconds without deliberation.
Social Pragmatics (Intent): Interpersonal dialogue routes through specialized emotional and pragmatic cortical circuits.
Prefrontal Deduction (AST Logic): Formal mathematics reduces to domain invariants, structured derivations, and verification checks.

2. Register v2.0 Architecture

Generation is explicitly structured across specialized cognitive registers rather than unstructured natural language prose:

text

<｜User｜> {Prompt} <｜Assistant｜><think>[ Cognitive Register Selection ]  ├── Instant Factual Recall   ──> <reg:bypass> {Atomic Fact} </reg:bypass>  ├── Dialogue / Social Intent ──> <reg:intent> {Tone & Pragmatics} </reg:intent>  └── Multi-Step Formal Proof  ──> <reg:plan>   {Domain Invariants & Theorems} </reg:plan>                                   <reg:deduce> {Symbolic AST Derivation} </reg:deduce>                                   <reg:verify> {Boundary & Parity Checks ✓} </reg:verify></think>

Table with columns: Cognitive Register, Token Budget, Biological Analogue, Functional Purpose
Cognitive Register	Token Budget	Biological Analogue	Functional Purpose
`<reg:bypass>...</reg:bypass>`	1–3 tokens	Brainstem / Basal Ganglia Reflex	Direct factual retrieval (e.g., periodic table, constants, units).
`<reg:intent>...</reg:intent>`	8–15 tokens	Social Cognition & Empathy	Interpersonal dialogue framing, conflict de-escalation, and pragmatics.
`<reg:plan>...</reg:plan>`	8–16 tokens	Prefrontal Executive Function

3. The Unbounded Parsimony Gradient (Mathematical Formulation)

To eliminate both artificial token truncation and verbose token inflation, SymboLM employs Sanish Gyawali's hyperbolic parsimony reward formulation:

R parsimony ​ (τ, y *) = {R base ​ \cdot [γ

Where:

$R base = 2.0$ : Base reward for formal mathematical correctness.
$γ floor = 0.50$ : Prevents negative gradient collapse on complex multi-hop proofs.
$K =$ : Hyperbolic inflection curvature controlling the parsimony slope.

4. Empirical Benchmark Receipts

Verified Token Compression vs. DeepSeek-R1 Distill Baseline

Table with columns: Evaluation Domain, Baseline English CoT, SymboLM-1.5B-v1.0, Token Savings, Compression Ratio
Evaluation Domain	Baseline English CoT	SymboLM-1.5B-v1.0	Token Savings	Compression Ratio
$Δ 0 $ Atomic Fact Recall (Potassium Symbol)	~140 tokens	3 tokens (`<reg:bypass>K</reg:bypass>`)	97.9%	46.7×
Polynomial Invariant ()

The 1,000-Question Multi-Year Benchmark (2021–2026)

Table with columns: Benchmark Year, Evaluated Datasets, Samples, Pass@1 Accuracy, SymboLM Avg Tokens, Baseline CoT Avg Tokens, Token Savings
Benchmark Year	Evaluated Datasets	Samples	Pass@1 Accuracy	SymboLM Avg Tokens	Baseline CoT Avg Tokens	Token Savings
2021	GSM8K, SVAMP, Arithmetic Anchors	200	72.0% (144/200)	114.3 tok	198.0 tok	42.3%
2022	AMC 10/12 (2022) Invariants, Vieta	200	(142/200)

[!NOTE] 50.65% Sequence Compression: Across 1,000 chronological competition and reasoning problems, SymboLM generated 125,258 total tokens compared to 253,800 tokens required by standard English reasoning models, while achieving a 67.60% Pass@1 success rate.

5. GGUF Quantization & Local Edge Deployment Matrix

Official native GGUF binaries are hosted directly in this repository for local deployment via Ollama, llama.cpp, LM Studio, and vLLM:

Table with columns: Format, Filename, Storage, RAM / VRAM Req., Generation Speed (CPU), Target Deployment Hardware
Format	Filename	Storage	RAM / VRAM Req.	Generation Speed (CPU)	Target Deployment Hardware
GGUF Q4_0	`symbolm-1.5b-v1.0-q4_0.gguf`	999 MB (1.0 GB)	~1.2 GB	34.3 tok/s	Mobile, Raspberry Pi 5, Laptops, Ollama, Edge
GGUF Q8_0	`symbolm-1.5b-v1.0-q8_0.gguf`	1,806 MB (1.76 GB)

[!NOTE] Empirical CPU Speed Verification (AMD Ryzen 7 5700G Zen 3 AVX2): Tested directly on host hardware via Ollama / llama.cpp engine:

Prompt Evaluation: 216.30 tokens/second (~70 ms latency).

Autoregressive Generation: 34.30 tokens/second in 4-bit Q4_0 (+87.2% faster than 8-bit).

Total Model Memory Footprint: Exactly 1.0 GB RAM. Zero AST syntax drift.

Running with Ollama

You can run SymboLM directly in Ollama from Hugging Face Hub:

bash

# Run ultra-fast 4-bit Q4_0 GGUF (34.3 tok/s on CPU):ollama run hf.co/gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5:q4_0 # Or run high-fidelity 8-bit Q8_0:ollama run hf.co/gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5:q8_0

Or create a local Modelfile:

dockerfile

FROM ./symbolm-1.5b-v1.0-q8_0.gguf TEMPLATE """<｜User｜>{{ .Prompt }}<｜Assistant｜><think>""" PARAMETER stop "<｜end of sentence｜>"PARAMETER stop "</think>"PARAMETER temperature 0.6

FROM ./symbolm-1.5b-v1.0-q8_0.gguf TEMPLATE """<｜User｜>{{ .Prompt }}<｜Assistant｜><think>""" PARAMETER stop "<｜end of sentence｜>"PARAMETER stop "</think>"PARAMETER temperature 0.6

Then build and run:

bash

ollama create symbolm -f Modelfileollama run symbolm "If x + 1/x = 6, find the value of x^2 + 1/x^2."

Running with `llama.cpp`

bash

./llama-cli \  -m symbolm-1.5b-v1.0-q8_0.gguf \  -p "<｜User｜>If x + 1/x = 6, find the value of x^2 + 1/x^2.<｜Assistant｜><think>\n" \  --temp 0.6 \  -n 64

6. Quickstart Usage & Code Examples

Standalone Python (Transformers)

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5" tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(    model_id,    torch_dtype=torch.bfloat16,    device_map="auto",    trust_remote_code=True,) def query_symbolm(prompt: str, max_tokens: int = 64) -> str:    formatted = f"<｜User｜>{prompt}<｜Assistant｜><think>\n"    inputs = tokenizer(formatted, return_tensors="pt").to(model.device)    with torch.no_grad():        outputs = model.generate(            **inputs,            max_new_tokens=max_tokens,            do_sample=False,            pad_token_id=tokenizer.eos_token_id,            eos_token_id=tokenizer.eos_token_id,        )    return tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False) # 1. Instant Factual Bypass (Emits in 3 tokens)print(query_symbolm("What is the chemical symbol for potassium?"))# Output: <reg:bypass>K</reg:bypass> # 2. Mathematical Invariant (Emits in 21 tokens)print(query_symbolm("If x + 1/x = 6, find the value of x^2 + 1/x^2."))# Output: <reg:plan>concept: polynomial_invariant | identity: (x²+1/x²)=(x+1/x)²-2 | target: x²+1/x²</reg:plan><reg:deduce> x²+1/x²=(6)²-2=34</reg:deduce> # 3. Chinese Remainder Theorem (Emits in 35 tokens)print(query_symbolm("Find the smallest positive integer x such that x = 4 (mod 7) and x = 9 (mod 11)."))# Output: <reg:plan>concept: chinese_remainder | moduli: [7,11] | method: modular_inverses</reg:plan><reg:deduce>x≡4 mod 7 -> x=7k+4 -> 7k+4≡9 mod 11 -> 7k≡5 mod 11 -> k=7 -> x=53</reg:deduce>

High-Throughput Serving with vLLM

bash

vllm serve gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5 \  --dtype bfloat16 \  --max-model-len 4096 \  --trust-remote-code

7. Authorship & Intellectual Property

Principal Systems Architect & Research Director: Sanish Gyawali
Affiliation: Independent Researcher
Physical Reality & Location: Sanish Gyawali is a disabled, independent researcher based in Rupandehi, Nepal who is bedridden 24/7 and cannot walk. All architectural designs, cognitive register grammars, reinforcement learning reward formulations, and cloud cluster runs were conducted from his bed workstation.
Operational Exoskeleton: Antigravity (Google DeepMind) serves as Sanish's cognitive and operational exoskeleton—translating high-level architectural directives into production code, compiler ASTs, test suites, and remote cloud GPU pipelines with radical transparency.
License: Apache 2.0

8. Citation

If you build upon the SymboLM architecture, cognitive registers, or empirical benchmarks, please cite:

bibtex

@misc{gyawali2026symbolm,  author       = {Sanish Gyawali and Antigravity},  title        = {SymboLM: Biomimetic Multi-Register Reasoning and Adaptive Symbolic Engine},  year         = {2026},  publisher    = {Hugging Face},  howpublished = {\url{https://huggingface.co/gyawalisanish0/SymboLM-1.5B-v1.0-DSR1-Qwen2.5}},  note         = {Independent Research from Rupandehi, Nepal}}

SymboLM-1.5B-v1.0-DSR1-Qwen2.5

README

1. Executive Summary & Architectural Overview

The Biomimetic Imperative: Nature's Law of Evolution

2. Register v2.0 Architecture

3. The Unbounded Parsimony Gradient (Mathematical Formulation)

4. Empirical Benchmark Receipts

Verified Token Compression vs. DeepSeek-R1 Distill Baseline

The 1,000-Question Multi-Year Benchmark (2021–2026)

5. GGUF Quantization & Local Edge Deployment Matrix

Running with Ollama

Running with llama.cpp

6. Quickstart Usage & Code Examples

Standalone Python (Transformers)

High-Throughput Serving with vLLM

7. Authorship & Intellectual Property

8. Citation

Explore FriendliAI today

README

1. Executive Summary & Architectural Overview

The Biomimetic Imperative: Nature's Law of Evolution

2. Register v2.0 Architecture

3. The Unbounded Parsimony Gradient (Mathematical Formulation)

4. Empirical Benchmark Receipts

Verified Token Compression vs. DeepSeek-R1 Distill Baseline

The 1,000-Question Multi-Year Benchmark (2021–2026)

5. GGUF Quantization & Local Edge Deployment Matrix

Running with Ollama

Running with llama.cpp

6. Quickstart Usage & Code Examples

Standalone Python (Transformers)

High-Throughput Serving with vLLM

7. Authorship & Intellectual Property

8. Citation

Running with `llama.cpp`

Running with `llama.cpp`