初始化项目，由ModelHub XC社区提供模型
Model: Xerv-AI/MAXWELL Source: Original Platform
This commit is contained in:
ModelHub XC
219
README.md
Normal file
219
README.md
Normal file
@@ -0,0 +1,219 @@
|
||||
|
||||
---
|
||||
language:
|
||||
- en
|
||||
license: apache-2.0
|
||||
base_model: unsloth/Qwen2.5-Math-1.5B-Instruct-bnb-4bit
|
||||
tags:
|
||||
- stem
|
||||
- mathematics
|
||||
- physics
|
||||
- unsloth
|
||||
- qwen2.5-math
|
||||
- reasoning
|
||||
- stss-framework
|
||||
- logic
|
||||
- analytical
|
||||
- science
|
||||
- meta-aggregation
|
||||
- 4bit
|
||||
- merged-f16
|
||||
library_name: transformers
|
||||
datasets:
|
||||
- Xerv-AI/TART
|
||||
metrics:
|
||||
- accuracy
|
||||
- math_verify
|
||||
model_creator: Xerv-AI
|
||||
model_name: MAXWELL
|
||||
pipeline_tag: text-generation
|
||||
|
||||
hf_space: Xerv-AI/Maxwell
|
||||
|
||||
# Leaderboard & Benchmark Specifications
|
||||
model-index:
|
||||
- name: MAXWELL (Qwen2.5-Math-1.5B-Instruct-STSS)
|
||||
results:
|
||||
- task:
|
||||
type: text-generation
|
||||
name: Grade School Mathematics
|
||||
dataset:
|
||||
name: GSM8K
|
||||
type: gsm8k
|
||||
split: test
|
||||
metrics:
|
||||
- type: accuracy
|
||||
value: 70.0
|
||||
name: Exact Match (Zero-Shot)
|
||||
- task:
|
||||
type: text-generation
|
||||
name: Competition Mathematics
|
||||
dataset:
|
||||
name: MATH-Hard
|
||||
type: lighteval/MATH-Hard
|
||||
config: default
|
||||
split: test
|
||||
metrics:
|
||||
- type: accuracy
|
||||
value: 60.0
|
||||
name: Exact Match (Boxed)
|
||||
- task:
|
||||
type: text-generation
|
||||
name: Professional Knowledge
|
||||
dataset:
|
||||
name: MMLU-Pro
|
||||
type: TIGER-Lab/MMLU-Pro
|
||||
config: default
|
||||
split: test
|
||||
metrics:
|
||||
- type: accuracy
|
||||
value: 45.0
|
||||
name: Multiple Choice Accuracy
|
||||
- task:
|
||||
type: text-generation
|
||||
name: Invitational Math
|
||||
dataset:
|
||||
name: AIME 2026
|
||||
type: MathArena/aime_2026
|
||||
split: train
|
||||
metrics:
|
||||
- type: accuracy
|
||||
value: 10.0
|
||||
name: Accuracy
|
||||
- task:
|
||||
type: text-generation
|
||||
name: Advanced Graduate Reasoning
|
||||
dataset:
|
||||
name: Humanity's Last Exam
|
||||
type: cais/hle
|
||||
config: default
|
||||
split: test
|
||||
metrics:
|
||||
- type: accuracy
|
||||
value: 0.0
|
||||
name: Exact String Match
|
||||
|
||||
# Technical Architecture Settings
|
||||
model_type: qwen2
|
||||
quantization: 4-bit (bitsandbytes)
|
||||
merged_format: fp16
|
||||
inference_framework:
|
||||
name: STSS (Systematic Temperature-Sweep Synthesis)
|
||||
phases:
|
||||
- generation_sweep: [0.1, 0.3, 0.5, 0.7, 0.9]
|
||||
- aggregation_method: neural_synthesis
|
||||
- logic_anchor: triboelectric_induction_verification
|
||||
max_position_embeddings: 4096
|
||||
rope_scaling:
|
||||
type: linear
|
||||
factor: 2.0
|
||||
|
||||
# Deployment Hardware
|
||||
hardware_specification:
|
||||
gpu: Tesla T4
|
||||
vram: 16GB
|
||||
optimization: Unsloth-Fast-Inference
|
||||
|
||||
---
|
||||
|
||||
|
||||
# MAXWELL: Model Card
|
||||
This document provides the technical specifications, training methodologies, and inference architecture for the MAXWELL model. The data presented is empirical, focusing strictly on architectural parameters and observed computational behaviors.
|
||||
## 1. Model Details
|
||||
### 1.1 Overview
|
||||
MAXWELL is a fine-tuned, specialized variant of the Qwen2.5-Math-1.5B-Instruct architecture. It is optimized for high-precision analytical reasoning, mathematical computation, and physics problem-solving. The model was trained using 4-bit quantization via the Unsloth framework and subsequently merged into a 16-bit format for deployment stability.
|
||||
### 1.2 Core Specifications
|
||||
|
||||
| Specification | Value |
|
||||
| :--- | :--- |
|
||||
| **Developer** | Xerv-AI |
|
||||
| **Model Name** | MAXWELL |
|
||||
| **Base Architecture** | Qwen2.5-Math-1.5B-Instruct |
|
||||
| **Parameter Count** | ~1.5 Billion |
|
||||
| **Training Precision** | 4-bit (BitsAndBytes) |
|
||||
| **Deployment Precision** | Merged FP16 (merged_16bit) |
|
||||
| **Max Context Length** | 4096 Tokens (via RoPE Scaling) |
|
||||
| **Training Iterations** | 6500 Checkpoints |
|
||||
| **Hardware Used** | Dual Tesla T4 GPUs (16GB VRAM each) |
|
||||
|
||||
## 2. Inference Architecture: STSS
|
||||
MAXWELL is uniquely designed to operate within a custom inference framework defined as **Systematic Temperature-Sweep Synthesis (STSS)**. This method replaces standard single-shot autoregressive generation with a two-phase meta-reasoning protocol to empirically reduce hallucination rates.
|
||||
### 2.1 Phase I: Spectrum Generation
|
||||
Instead of sampling at a fixed temperature, the framework forces the model to generate a set of candidate responses \mathcal{S} across a defined temperature grid G_\tau:
|
||||
* **Low Entropy (T \in [0.1, 0.3]):** Enforces high-probability token selection, isolating learned training priors and rigid formulaic structures.
|
||||
* **High Entropy (T \in [0.7, 0.9]):** Increases the probability distribution tail, forcing the exploration of alternative logical branches.
|
||||
### 2.2 Phase II: Neural Aggregation
|
||||
The model is re-prompted using the entire generated set \mathcal{S} as its context window. It acts as an aggregator function f_{agg} to synthesize the final output R_{final}:
|
||||
This aggregation is explicitly executed at T=0.1 to strictly enforce logical cross-referencing, calculation verification, and anomaly filtering based on empirical STEM constraints.
|
||||
## 3. Empirical Performance Observations
|
||||
Based on inference testing logs, the model exhibits the following data-driven characteristics:
|
||||
* **Pattern-Recognition Override:** In cognitive reflection tests (e.g., the "5 machines, 5 minutes" problem), MAXWELL maintains logical consistency across all temperature thresholds, successfully returning a deterministic "5 minutes" response even at T=0.9.
|
||||
* **Triboelectric Physics Accuracy:** Requires explicit anchoring prompts during aggregation to override common dataset biases regarding electrostatic charge polarities (e.g., explicitly defining Glass + Silk = Positive).
|
||||
* **Zero-Shot Consensus:** When presented with non-complex strings (e.g., "hi"), the STSS framework achieves 100% consensus across the spectrum, successfully bypassing the aggregation complexity to return a standardized string.
|
||||
## 4. Limitations & Computational Overhead
|
||||
### 4.1 Token Saturation
|
||||
Because the STSS framework requires injecting five complete reasoning paths into the Phase II prompt, long-form calculus or multi-step proofs will trigger a context truncation limit. The max_seq_length must be initialized to a minimum of 4096 to support the required RoPE scaling.
|
||||
### 4.2 Compute Multiplier
|
||||
Standard LLM inference processes one generation pass. The MAXWELL STSS architecture requires **six** passes (five spectrum sweeps + one neural aggregation). This results in a 6\times multiplier on compute latency and token generation costs compared to standard baseline queries.
|
||||
## 5. Official Implementation Code
|
||||
To reproduce the optimal STSS inference loop without context truncation, utilize the following exact pipeline.
|
||||
```python
|
||||
from unsloth import FastLanguageModel
|
||||
from transformers import TextStreamer
|
||||
import torch
|
||||
# Configuration
|
||||
MODEL_NAME = "Xerv-AI/MAXWELL"
|
||||
MAX_CONTEXT = 4096
|
||||
# Load Base
|
||||
model, tokenizer = FastLanguageModel.from_pretrained(
|
||||
model_name = MODEL_NAME,
|
||||
max_seq_length = MAX_CONTEXT,
|
||||
load_in_4bit = True,
|
||||
)
|
||||
FastLanguageModel.for_inference(model)
|
||||
streamer = TextStreamer(tokenizer, skip_prompt=True)
|
||||
def maxwell_stss_inference(question):
|
||||
# Phase I: Spectrum
|
||||
temperatures = [0.1, 0.3, 0.5, 0.7, 0.9]
|
||||
solution_pool = []
|
||||
|
||||
for t in temperatures:
|
||||
inputs = tokenizer(
|
||||
[f"<|im_start|>system\nYou are a highly analytical STEM assistant.<|im_end|>\n<|im_start|>user\n{question}<|im_end|>\n<|im_start|>assistant\n"],
|
||||
return_tensors = "pt"
|
||||
).to("cuda")
|
||||
|
||||
output = model.generate(
|
||||
**inputs,
|
||||
max_new_tokens=450,
|
||||
temperature=t,
|
||||
use_cache=True
|
||||
)
|
||||
decoded = tokenizer.batch_decode(output)[0].split("<|im_start|>assistant\n")[-1].replace("<|im_end|>", "").strip()
|
||||
solution_pool.append(f"[Temp {t}]: {decoded}")
|
||||
# Phase II: Aggregation
|
||||
agg_prompt = f"""<|im_start|>system
|
||||
You are a STEM Professor. Compare the 5 solutions below.
|
||||
Even if they all agree, you must:
|
||||
1. Explain WHY the consensus is correct.
|
||||
2. Formulate a final, perfect response using LaTeX.
|
||||
<|im_end|>
|
||||
<|im_start|>user
|
||||
PROBLEM: {question}
|
||||
SOLUTIONS:
|
||||
{chr(10).join(solution_pool)}
|
||||
<|im_end|>
|
||||
<|im_start|>assistant
|
||||
<reasoning>
|
||||
Based on the provided candidates, there is a 100% consensus. Here is the final verification:"""
|
||||
final_inputs = tokenizer([agg_prompt], return_tensors="pt").to("cuda")
|
||||
|
||||
final_output = model.generate(
|
||||
**final_inputs,
|
||||
max_new_tokens=1024,
|
||||
temperature=0.1,
|
||||
streamer=streamer,
|
||||
use_cache=True
|
||||
)
|
||||
return "Generation Complete."
|
||||
```
|
||||
Reference in New Issue
Block a user