Instructions to use ramankrishna10/npc-reason with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use ramankrishna10/npc-reason with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf ramankrishna10/npc-reason:Q4_K_M # Run inference directly in the terminal: llama cli -hf ramankrishna10/npc-reason:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf ramankrishna10/npc-reason:Q4_K_M # Run inference directly in the terminal: llama cli -hf ramankrishna10/npc-reason:Q4_K_M
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf ramankrishna10/npc-reason:Q4_K_M # Run inference directly in the terminal: ./llama-cli -hf ramankrishna10/npc-reason:Q4_K_M
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf ramankrishna10/npc-reason:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama-cli -hf ramankrishna10/npc-reason:Q4_K_M
Use Docker
docker model run hf.co/ramankrishna10/npc-reason:Q4_K_M
- LM Studio
- Jan
- vLLM
How to use ramankrishna10/npc-reason with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "ramankrishna10/npc-reason" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "ramankrishna10/npc-reason", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/ramankrishna10/npc-reason:Q4_K_M
- Ollama
How to use ramankrishna10/npc-reason with Ollama:
ollama run hf.co/ramankrishna10/npc-reason:Q4_K_M
- Unsloth Studio
How to use ramankrishna10/npc-reason with Unsloth Studio:
Install Unsloth Studio (macOS, Linux, WSL)
curl -fsSL https://unsloth.ai/install.sh | sh # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for ramankrishna10/npc-reason to start chatting
Install Unsloth Studio (Windows)
irm https://unsloth.ai/install.ps1 | iex # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for ramankrishna10/npc-reason to start chatting
Using HuggingFace Spaces for Unsloth
# No setup required # Open https://huggingface.co/spaces/unsloth/studio in your browser # Search for ramankrishna10/npc-reason to start chatting
- Atomic Chat new
- Docker Model Runner
How to use ramankrishna10/npc-reason with Docker Model Runner:
docker model run hf.co/ramankrishna10/npc-reason:Q4_K_M
- Lemonade
How to use ramankrishna10/npc-reason with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull ramankrishna10/npc-reason:Q4_K_M
Run and chat with the model
lemonade run user.npc-reason-Q4_K_M
List all available models
lemonade list
File size: 11,415 Bytes
1ba301a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 | """NPC Reason — mechanical step verifier (THE core artifact).
PURE CODE. No model, no LLM judgment anywhere. This is what makes the
"verifiable-rate" metric un-fakeable, and it is reused verbatim as the RL reward
signal in a later dispatch — so it is frozen (VERIFIER.lock) the moment its tests pass.
DEFINITION (committed; see reports/PREREG.md and VERIFIER.lock):
- A reasoning chain is a sequence of steps ending in a final answer.
- A load-bearing numeric step carries an inline checkable assertion in the canonical
form <<EXPR = RESULT>> where EXPR is an arithmetic/algebraic expression over
numbers and earlier-BOUND variables, and RESULT is the claimed value.
- A step is VERIFIED if a SymPy evaluation of EXPR equals RESULT within tolerance
(exact for integers/rationals; 1e-6 relative for floats).
- A chain is VERIFIABLE iff:
(a) it has >=1 load-bearing <<...>> assertion (no bare asserted numbers drive it),
(b) every such assertion VERIFIES, and
(c) the final answer equals the RESULT of the last load-bearing step
(the chain COMPOSES to its conclusion).
- A chain is CORRECT iff its final answer equals the gold answer. CORRECT and
VERIFIABLE are INDEPENDENT axes.
TOLERANCE POLICY (explicit, frozen):
- Parse EXPR and RESULT with sympy.sympify (after a small, fixed normalization).
- Exact branch: if simplify(EXPR_value - RESULT_value) == 0 -> VERIFIED.
- Float branch: else, if both are finite numbers and
|EXPR_value - RESULT_value| <= 1e-6 * max(1, |RESULT_value|) -> VERIFIED.
- Otherwise NOT verified.
FAIL-CLOSED (conservative by design):
- Unparseable EXPR or RESULT -> NOT verified (reason recorded).
- EXPR references an UNBOUND symbol -> NOT verified (cannot confirm).
- Any exception during evaluation -> NOT verified.
A chain only counts VERIFIABLE if the checker can ACTUALLY confirm every load-bearing step.
VARIABLE BINDING (v1, documented):
- A step may bind a name: `let total = <<3*8 = 24>>` or `total = <<3*8 = 24>>`.
The name binds to the (parsed) RESULT value and is usable by later EXPRs.
- Binding tests internal consistency: each later step is checked against the values
the chain itself previously stated.
- V1 SCOPE: assertions are concrete-valued (EXPR evaluates to a number once prior
bindings are substituted). Algebra with FREE variables / equation-solving such as
`<<x**2 - 4 = 0>>` with x unbound is OUT OF SCOPE for v1 and FAILS CLOSED
(recorded reason: "unbound symbol"). It is noted as a verifier-v2 extension.
"""
from __future__ import annotations
import re
from dataclasses import dataclass, field
from typing import Optional
import sympy
from sympy import Rational, simplify, sympify
# ----------------------------------------------------------------------------- #
# Patterns
# ----------------------------------------------------------------------------- #
# Optional binding name, then << EXPR = RESULT >>. Non-greedy EXPR stops at the
# FIRST '=' inside the brackets. Binding name must start with a letter/underscore,
# so numeric tokens to the left (e.g. "2 + 2 =") are never captured as a name.
ASSERTION = re.compile(
r"(?:(?:let\s+)?([A-Za-z_]\w*)\s*=\s*)?<<\s*(.+?)\s*=\s*(.+?)\s*>>"
)
# Final-answer extractors, tried in priority order.
_BOXED = re.compile(r"\\boxed\{\s*([^{}]+?)\s*\}")
_GSM = re.compile(r"####\s*([^\n]+)")
_ANSWER_IS = re.compile(
r"(?:final answer|the answer is|answer\s*[:=])\s*\$?\\?\(?\s*"
r"([+-]?[\d.,/eE^*+\-() ]*\d)",
re.IGNORECASE,
)
# Characters/sequences to normalize before sympify.
_THOUSANDS = re.compile(r"(?<=\d),(?=\d{3}\b)")
_NORM_REPLACE = (
(r"\times", "*"),
(r"\cdot", "*"),
(r"\div", "/"),
(r"\left", ""),
(r"\right", ""),
("^", "**"),
("%", "/100"),
("$", ""),
("×", "*"),
("÷", "/"),
("−", "-"), # unicode minus
)
@dataclass
class StepResult:
expr: str
claimed: str
ok: bool
reason: str
binding: Optional[str] = None
@dataclass
class ChainRecord:
n_assertions: int = 0
n_verified: int = 0
has_loadbearing_assertions: bool = False
all_assertions_verified: bool = False
final_answer: Optional[str] = None
final_answer_value: Optional[object] = None
composes_to_final: bool = False
verifiable: bool = False
correct: Optional[bool] = None
verified_and_correct: Optional[bool] = None
steps: list = field(default_factory=list)
failures: list = field(default_factory=list)
def as_dict(self) -> dict:
return {
"n_assertions": self.n_assertions,
"n_verified": self.n_verified,
"has_loadbearing_assertions": self.has_loadbearing_assertions,
"all_assertions_verified": self.all_assertions_verified,
"final_answer": self.final_answer,
"composes_to_final": self.composes_to_final,
"verifiable": self.verifiable,
"correct": self.correct,
"verified_and_correct": self.verified_and_correct,
"failures": self.failures,
"steps": [
{"expr": s.expr, "claimed": s.claimed, "ok": s.ok,
"reason": s.reason, "binding": s.binding}
for s in self.steps
],
}
# ----------------------------------------------------------------------------- #
# Numeric core
# ----------------------------------------------------------------------------- #
def _normalize(raw: str) -> str:
s = raw.strip()
s = _THOUSANDS.sub("", s) # 1,234 -> 1234 (drop thousands separators)
for a, b in _NORM_REPLACE:
s = s.replace(a, b)
# \frac{a}{b} -> ((a)/(b))
s = re.sub(r"\\d?frac\{([^{}]+)\}\{([^{}]+)\}", r"((\1)/(\2))", s)
s = s.replace("\\", " ")
return s.strip()
def _to_sympy(raw: str, bindings: dict):
"""sympify a normalized token; raises on failure (caller fails closed)."""
expr = sympify(_normalize(raw), locals=bindings, rational=True)
return expr
def _values_match(a, b) -> bool:
"""Exact for rationals/integers; 1e-6 relative for floats."""
diff = simplify(a - b)
if diff == 0:
return True
try:
if diff.free_symbols:
return False
fa, fb = float(a), float(b)
return abs(fa - fb) <= 1e-6 * max(1.0, abs(fb))
except (TypeError, ValueError):
return False
def verify_assertion(expr: str, claimed: str, bindings: dict) -> StepResult:
"""Evaluate one <<EXPR = RESULT>> against current bindings. Fail-closed."""
try:
e = _to_sympy(expr, bindings)
except Exception as ex: # noqa: BLE001 — fail closed on any parse error
return StepResult(expr, claimed, False, f"unparseable expr: {ex}")
try:
c = _to_sympy(claimed, bindings)
except Exception as ex: # noqa: BLE001
return StepResult(expr, claimed, False, f"unparseable result: {ex}")
# EXPR must reduce to a concrete value (v1 scope: no free variables).
if getattr(e, "free_symbols", set()):
unbound = ", ".join(sorted(str(s) for s in e.free_symbols))
return StepResult(expr, claimed, False, f"unbound symbol(s): {unbound}")
try:
ok = _values_match(e, c)
except Exception as ex: # noqa: BLE001
return StepResult(expr, claimed, False, f"comparison error: {ex}")
reason = "verified" if ok else f"mismatch: {expr} -> {e} != {claimed}"
return StepResult(expr, claimed, ok, reason)
# ----------------------------------------------------------------------------- #
# Final answer
# ----------------------------------------------------------------------------- #
def extract_final_answer(text: str) -> Optional[str]:
"""Last \\boxed{}, else last ####, else 'the answer is X'. None if absent."""
boxed = _BOXED.findall(text)
if boxed:
return boxed[-1].strip()
gsm = _GSM.findall(text)
if gsm:
return gsm[-1].strip()
m = list(_ANSWER_IS.finditer(text))
if m:
return m[-1].group(1).strip().rstrip(".")
return None
def _safe_value(raw: str, bindings: dict):
try:
v = _to_sympy(raw, bindings)
return None if getattr(v, "free_symbols", set()) else v
except Exception: # noqa: BLE001
return None
# ----------------------------------------------------------------------------- #
# Chain
# ----------------------------------------------------------------------------- #
def verify_chain(text: str, gold_answer=None) -> dict:
"""Mechanically derive every field. No judgment. Returns ChainRecord.as_dict()."""
rec = ChainRecord()
bindings: dict = {}
for m in ASSERTION.finditer(text):
name, expr, claimed = m.group(1), m.group(2), m.group(3)
res = verify_assertion(expr, claimed, bindings)
res.binding = name
rec.steps.append(res)
rec.n_assertions += 1
if res.ok:
rec.n_verified += 1
else:
rec.failures.append({"expr": expr, "claimed": claimed, "reason": res.reason})
# Bind the name to the CLAIMED result value (internal-consistency semantics),
# whenever the result parses to a concrete value — even if the step failed,
# so downstream reasons are about the downstream step, not a cascade.
if name:
v = _safe_value(claimed, bindings)
if v is not None:
bindings[name] = v
rec.has_loadbearing_assertions = rec.n_assertions > 0
rec.all_assertions_verified = (
rec.has_loadbearing_assertions and rec.n_verified == rec.n_assertions
)
# Final answer + composition.
fa = extract_final_answer(text)
rec.final_answer = fa
fa_val = _safe_value(fa, bindings) if fa is not None else None
rec.final_answer_value = fa_val
if rec.has_loadbearing_assertions and fa_val is not None:
last_claimed = rec.steps[-1].claimed
last_val = _safe_value(last_claimed, bindings)
if last_val is not None:
try:
rec.composes_to_final = _values_match(fa_val, last_val)
except Exception: # noqa: BLE001
rec.composes_to_final = False
if not rec.composes_to_final and rec.has_loadbearing_assertions:
rec.failures.append({"reason": "final answer does not compose from last step"})
rec.verifiable = (
rec.has_loadbearing_assertions
and rec.all_assertions_verified
and rec.composes_to_final
)
# Correctness (independent axis).
if gold_answer is not None:
gold_val = _safe_value(str(gold_answer), {})
if fa_val is not None and gold_val is not None:
try:
rec.correct = _values_match(fa_val, gold_val)
except Exception: # noqa: BLE001
rec.correct = False
else:
# Fall back to normalized string compare (handles non-numeric MATH answers).
rec.correct = (
fa is not None
and _normalize(fa).replace(" ", "") == _normalize(str(gold_answer)).replace(" ", "")
)
rec.verified_and_correct = bool(rec.verifiable and rec.correct)
return rec.as_dict()
if __name__ == "__main__":
import json
import sys
blob = sys.stdin.read()
print(json.dumps(verify_chain(blob), indent=2, default=str))
|