Text Generation
Safetensors
English
llama
conversational
Eval Results (legacy)
Wildalign-350M-base / README.md
WildMann's picture
Update README.md
bb38e98 verified
|
Raw
History Blame Contribute Delete
7.95 kB
metadata
pipeline_tag: text-generation
license: apache-2.0
datasets:
  - HuggingFaceTB/finemath
  - HuggingFaceFW/finepdfs_edu_50BT-dclm_30BT-fineweb_edu_20BT-shuffled
language:
  - en
model-index:
  - name: Wildalign-350M-base
    results:
      - task:
          type: arc_challenge
          name: ARC Challenge
        dataset:
          type: arc_challenge
          name: ARC Challenge
          version: 1
        metrics:
          - name: acc
            type: acc
            value: 0.2082
            stderr: ± 0.0119
            verified: false
          - name: acc_norm
            type: acc_norm
            value: 0.2543
            stderr: ± 0.0127
            verified: false
      - task:
          type: arc_easy
          name: ARC Easy
        dataset:
          type: arc_easy
          name: ARC Easy
          version: 1
        metrics:
          - name: acc
            type: acc
            value: 0.5442
            stderr: ± 0.0102
            verified: false
          - name: acc_norm
            type: acc_norm
            value: 0.4701
            stderr: ± 0.0102
            verified: false
      - task:
          type: boolq
          name: BoolQ
        dataset:
          type: boolq
          name: BoolQ
          version: 2
        metrics:
          - name: acc
            type: acc
            value: 0.6049
            stderr: ± 0.0086
            verified: false
      - task:
          type: hellaswag
          name: HellaSwag
        dataset:
          type: hellaswag
          name: HellaSwag
          version: 1
        metrics:
          - name: acc
            type: acc
            value: 0.2984
            stderr: ± 0.0046
            verified: false
          - name: acc_norm
            type: acc_norm
            value: 0.3292
            stderr: ± 0.0047
            verified: false
      - task:
          type: piqa
          name: PIQA
        dataset:
          type: piqa
          name: PIQA
          version: 1
        metrics:
          - name: acc
            type: acc
            value: 0.6496
            stderr: ± 0.0111
            verified: false
          - name: acc_norm
            type: acc_norm
            value: 0.6311
            stderr: ± 0.0113
            verified: false
      - task:
          type: truthfulqa_mc2
          name: TruthfulQA MC2
        dataset:
          type: truthfulqa_mc2
          name: TruthfulQA MC2
          version: 3
        metrics:
          - name: acc
            type: acc
            value: 0.4118
            stderr: ± 0.0148
            verified: false
      - task:
          type: winogrande
          name: Winogrande
        dataset:
          type: winogrande
          name: Winogrande
          version: 1
        metrics:
          - name: acc
            type: acc
            value: 0.5233
            stderr: ± 0.0140
            verified: false

🐺 Wildalign-350M-base

Experimental Exotic Autoregressive Language Model

This repository hosts Wildalign‑350M‑base—an experimental, lightweight autoregressive language model designed to investigate the impact of high-quality "exotic" textual data on model alignment at the GPT-2 base parameter scale.

⚠️ Critical Content Warning The model's training corpus intentionally incorporates high-quality "exotic" textual data. Consequently, under specific prompt guidance, the model may generate content that falls under NSFW categories, including but not limited to mature/non-mainstream narratives, roleplay, or thematic scenarios. While the model remains safe for standard use cases, we advise caution and appropriate content filtering for public deployment.


🧠 Core Model Specifications

Attribute Details
Architecture Autoregressive Decoder-Only
Parameter Scale ~350 Million (Experimental small-scale)
Training Hardware Single NVIDIA H100 GPU
Training Duration ~110 Hours
Training Precision bfloat16 (bf16)
Max Sequence Length 1,024 Tokens

📚 Training Corpus Breakdown (Total: ~11B Tokens)

The dataset was deliberately partitioned into three distinct components to isolate the impact of the exotic data on general alignment:

  1. Public Mixed Educational/General Corpus (10B tokens)
    • Source: HuggingFaceFW/finepdfs_edu_50BT-dclm_30BT-fineweb_edu_20BT-shuffled
    • Purpose: Serves as the foundational, high-quality mainstream baseline.
  2. Mathematics Specialized Corpus (~1B tokens)
    • Source: HuggingFaceTB/finemath
    • Purpose: Injects structured, logical reasoning capabilities.
  3. High-Quality Exotic Text Corpus (0.3B tokens)
    • Processing: Strictly curated and distilled from 20GB of raw data down to 1GB of high-density content.
    • Content Type: Well-crafted narratives, storytelling, and complex roleplay scenarios.

🎯 Research Objectives & Preliminary Findings

Primary Research Question

Does the inclusion of high-quality "exotic" or mature story-based textual data fundamentally derail or negatively impact the alignment of small-scale language models?

Preliminary Insights (Measured at 3B token checkpoint)

  1. No Alignment Degradation: We observed no statistically significant negative impact on core model alignment resulting from the inclusion of exotic data.
  2. Architectural Compatibility: Exotic long-form narrative content is not intrinsically problematic for long-attention architectures, though practical deployment still heavily relies on robust content moderation.
  3. Scalability Proof-of-Concept: Achieving these stable results on a small-scale model provides a valid, resource-efficient proof-of-concept for larger-scale alignment research.

📈 Baseline Evaluation Results

Note: This model was explicitly developed for alignment research rather than zero-shot benchmark dominance. Zero-shot accuracy was not optimized for, resulting in expected baseline performance metrics.

Task Version Filter n-shot Metric Value Stderr
arc_challenge 1 none 0 acc
acc_norm
0.2082
0.2543
± 0.0119
± 0.0127
arc_easy 1 none 0 acc
acc_norm
0.5442
0.4701
± 0.0102
± 0.0102
boolq 2 none 0 acc 0.6049 ± 0.0086
hellaswag 1 none 0 acc
acc_norm
0.2984
0.3292
± 0.0046
± 0.0047
piqa 1 none 0 acc
acc_norm
0.6496
0.6311
± 0.0111
± 0.0113
truthfulqa_mc2 3 none 0 acc 0.4118 ± 0.0148
winogrande 1 none 0 acc 0.5233 ± 0.0140

📋 Nomenclature

The name "Wildalign" is a portmanteau of "Wild/Exotic" (referencing our deliberate use of non-mainstream data components) and "Alignment" (the core research focus). The moniker reflects our primary study goal: navigating the wild to determine if mixing mainstream rules with unusual data sources disrupts model safety and coherence.


🤝 Support & Collaboration

If you find this research valuable and wish to help fuel further development, your support is deeply appreciated!

Support me on Ko-fi

All donations received will be strictly allocated to the computational costs of large-scale pre-training—driving research, model optimization, and scaling efforts for more generalized AI systems.

💻 Call for Compute Collaborators We warmly welcome anyone with dormant GPU/compute budgets to join our community. Your computational resources will play a vital role in advancing these foundation model projects, allowing us to accelerate the pace of open-source AI development together. Reach out if you're interested in collaborating!