NOSIBLE/prediction-v1.1-base

Prediction v1.1 Base

Changelog

• v1.0.0: Initial version

prediction-v1.1-base is a next-generation classification model designed to determine whether a short text snippet contains a prediction or not. For example,

Contains a prediction:

Google will increase the development of data centers in 2026 by 25%.

Does not contain a prediction:

Google increased the development of data centers in 2025 by 25%.

This model is fine-tuned from Qwen3-0.6B. Our training corpus consists of 100k real-world search results sourced from the Nosible Search Feeds product, giving the model broad exposure to real, noisy, and highly varied financial text as it appears on the web. You can find the open-source dataset here.

Why this model matters

1. Accurate detection of predictive statements Specialized training allows it to reliably distinguish predictive intent from descriptive or historical text, even when cues are subtle.

2. Robust on real-world financial and web data Because it’s trained on noisy, naturally occurring search-feed content, it handles messy, unstructured inputs without heavy preprocessing.

3. Scalable prediction mining Its lightweight architecture enables fast, large-scale extraction of forecasts, predictions, and estimates across massive text corpora.

Performance overview

prediction-v1.1-base consistently outperforms larger state-of-the-art LLMs while being deployable at a much lower cost. We computed the validation set accuracy based on a sample of 1,000 points from our validation set.

Cost per 1M tokens for the LLMs was calculated as a weighted average of input and output token costs using a 10:1 ratio (10× input cost + 1× output cost, divided by 11), based on pricing from OpenRouter. This reflects the ratio between our prompt used to label our dataset.

For the NOSIBLE model, we conservatively used the cost of Qwen-8B on OpenRouter with a 100:1 ratio since the model produces a single output token when used as described in this guide. Despite this, our model is still the cheapest option.

Class token mapping.

Because this is a classification model built off the Qwen3-0.6B, we mapped the prediction and not-prediction classes onto tokens. This is the mapping we chose.

{
    "prediction": "prediction",      # Contains a prediction.
    "not-prediction": "_prediction"  # Does not contain a prediction.
}

Strict Usage Requirements

1. Disable Thinking: You must set enable_thinking=False (or disable reasoning tokens).
2. Exact System Prompt: You must use the specific system prompt: "Classify whether it contains a prediction or does not contain a prediction."
3. Constrain Output: You must restrict generation to the valid labels (["prediction", "_prediction"]) using grammars, regex, or guided decoding.
   • SGLang: Use regex="(prediction|_prediction)" in the API call.
   • vLLM: Use guided_choice=["prediction", "_prediction"] in the API call.
   • llama.cpp / GGUF: Apply a GBNF grammar or regex to force selection from the list.
   • OpenAI / Structured Outputs: Use response_format or JSON Schema enforcement where supported.

Deviating from these requirements will severely impact performance and reliability.

Quickstart

Since this model was trained as a Causal LM using specific chat templates, you must use the apply_chat_template method with the specific system prompt used during training.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "NOSIBLE/prediction-v1.1-base"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    device_map="auto", 
    trust_remote_code=True, 
    torch_dtype=torch.bfloat16
)

# Define the input text
text = "The company is expected to record a profit margin of more than 15% next quarter."

# 1. Structure the prompt exactly as used in training
messages = [
    {"role": "system", "content": "Classify whether it contains a prediction or does not contain a prediction."},
    {"role": "user", "content": text},
]

# 2. Apply chat template
prompt = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True,
    enable_thinking=False # Must be set to false.
)

inputs = tokenizer([prompt], return_tensors="pt").to(model.device)

# 3. Generate the response (label)
# We limit max_new_tokens because only a single-word response is expected
outputs = model.generate(**inputs, max_new_tokens=1)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

# The model echoes the system and user messages in its output, so we extract the new text
print(response.split("<|im_start|>assistant\n")[-1])
# Expected Output: prediction

Deployment

For production deployment, you can use sglang>=0.4.6.post1 vllm>=0.8.5 to create an OpenAI compatible API endpoint.

The model is based on Qwen3-0.6B and as such can be deployed everywhere Qwen3-0.6B can. However, we recommend deploying using SGLang.

SGLang:

python -m sglang.launch_server --model-path Qwen/Qwen3-0.6B --reasoning-parser qwen3

Here is an example API call using an OpenAI compatible server to extract the probability for each label.

import math
from openai import OpenAI

# Initialize the client pointing to your vLLM server
client = OpenAI(
    base_url="http://localhost:8000/v1",  # Replace with your endpoint URL if remote
    api_key="EMPTY"
)

model_id = "NOSIBLE/prediction-v1.1-base"

# Input text to classify
text = "The company is expected to record a profit margin of more than 15% next quarter."

# Define the classification labels
labels = ["prediction", "_prediction"]

# Prepare the conversation
messages = [
    {"role": "system", "content": "Classify whether it contains a prediction or does not contain a prediction."},
    {"role": "user", "content": text},
]

# Make the API call
chat_completion = client.chat.completions.create(
    model=model_id,
    messages=messages,
    temperature=0,
    max_tokens=1,
    stream=False,
    logprobs=True,              # Enable log probabilities to calculate confidence
    top_logprobs=len(labels),   # Ensure we capture logprobs for our choices
    extra_body={
        "chat_template_kwargs": {"enable_thinking": False}, # Must be set to false.
        "regex": "(prediction|_prediction)",
    },
)

# Extract the response content
response_label = chat_completion.choices[0].message.content

# Extract the logprobs for the generated token to calculate confidence
first_token_logprobs = chat_completion.choices[0].logprobs.content[0].top_logprobs

print(f"--- Classification Results ---")
print(f"Input: {text}")
print(f"Predicted Label: {response_label}\n")

print("--- Label Confidence ---")
for lp in first_token_logprobs:
    # Convert log probability to percentage
    probability = math.exp(lp.logprob)
    print(f"Token: '{lp.token}' | Probability: {probability:.2%}")

Expected Output

--- Classification Results ---
Input: The company is expected to record a profit margin of more than 15% next quarter.
Predicted Label: prediction

--- Label Confidence ---
Token: 'prediction' | Probability: 99.98%
Token: '_prediction' | Probability: 0.02%

Using the regex restricts the output to one of prediction or _prediction, although substring matches may still occur.

Local Use: Applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers also support Qwen3 architectures.

Legal Notice: This model is a modification of the Qwen3-0.6B model. In compliance with the Apache 2.0 license, we retain all original copyright notices and provide this modification under the same license terms.

Training Details

Training Procedure

The model was fine-tuned using the Hugging Face Trainer with bf16 precision.

Preprocessing

• System Prompt: Added to every example.
• Masking: The user prompt and system instructions were masked (labels set to -100) so the model only calculates loss on the assistant's response (the label).
• Max Length: 2048 tokens.

Training Hyperparameters

Hyperparameter	Value
Learning Rate	2e-5
Scheduler	Cosine (Warmup ratio 0.03)
Batch Size	64
Epochs	2
Optimizer	AdamW Torch Fused
Precision	bfloat16
NEFTune Noise Alpha	5
Weight Decay	0.1

Limitations

While this model is optimized for efficiency and specific financial tasks, users should be aware of the following limitations:

• Parameter Size (0.6B): As a small language model, it lacks the deep reasoning capabilities of larger models (e.g., 7B or 70B parameters). It is designed for fast, specific classification tasks and may struggle with highly nuanced or ambiguous text that requires extensive world knowledge.
• Context Window: The model was trained with a maximum sequence length of 2048 tokens. For analyzing long financial documents (like 10-K filings or earnings call transcripts), the text must be chunked or truncated before processing.
• Domain Specificity: The model is primarily fine-tuned on financial contexts. It is not suitable for detecting predictions in other contexts (e.g., social media posts).
• Language Support: The model is trained primarily on English financial data. Its performance on non-English text or multi-lingual financial reports is not guaranteed.
• Factuality: This model analyzes the predictive nature of the text provided; it does not verify the factual accuracy of financial figures, dates, or claims within that text.

We utilized the following papers and datasets during the research and development of this model:

Papers:

• Qwen3: Qwen3 Technical Report. arXiv:2505.09388
• Qwen3 Guard: Qwen3Guard Technical Report. arXiv:2510.14276v1

Datasets:

• NOSIBLE Prediction: Nosible Ltd.

Disclaimer

• Not Financial Advice: The outputs of this model should not be interpreted as financial advice, investment recommendations, or an endorsement of any financial instrument or asset.
• Limitations: This model may not accurately classify predictions in highly complex, nuanced, or evolving financial contexts, including new market trends, highly specialized jargon, or sarcasm. Users are solely responsible for all decisions made based on its output.
• Risk: Financial markets are inherently volatile and risky. Never make investment decisions based solely on the output of an AI model. Always consult with a qualified financial professional.
• Task specific: This model only identifies whether a statement is prediction; it does not predict outcomes or evaluate likelihood.

Team & Credits

This model was developed and maintained by the following team:

• Matthew Dicks
• Simon van Dyk
• Gareth Warburton
• Stuart Reid

Citation

If you use this model, please cite it as follows:

@misc{nosible2025prediction,
  author = {NOSIBLE},
  title = {Prediction v1.1 Base},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Repository},
  howpublished = {https://huggingface.co/NOSIBLE/prediction-v1.1-base}
}