rsi-gpt-oss-20b

RSI - Recursive Self Improvement (LLM)

Model Card

We release open-weight metatune-gpt20b, fine tuned version of OpenAI's gpt-oss-20b model, this is one of the first public release recursive self improving AI.

• Generates new data for itself,
• Evaluates its performance, and
• Adjusts its own hyperparameters based on improvement metrics.

additional Model Information

Due to recursive self improvement method, there is no final model, but improved model, this is a multiple metacycle(generation) improved checkpoint model.

Use cases:

• math
• coding

Guardrails:

• generally, please set reasoning = "high", it will usually prevent jailbreaking and prompt injection
• use safety gpt oss 20b for guardrails before this model: openai/gpt-oss-safeguard-20b

Benchmark:

Inference examples

Transformers

You can use gpt-oss-120b and gpt-oss-20b with Transformers. If you use the Transformers chat template, it will automatically apply the harmony response format. If you use model.generate directly, you need to apply the harmony format manually using the chat template or use our openai-harmony package.

To get started, install the necessary dependencies to setup your environment:

pip install -U transformers kernels torch 

For Google Colab (free/Pro)

!pip install -q --upgrade torch

!pip install -q transformers triton==3.4 kernels

!pip uninstall -q torchvision torchaudio -y

Once, setup you can proceed to run the model by running the snippet below:

from transformers import pipeline
import torch
model_id = "EpistemeAI/rsi-gpt-oss-20b"
pipe = pipeline(
    "text-generation",
    model=model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Derive the Euler–Lagrange equation from the principle of stationary action."},
]

outputs = pipe(
    messages,
    max_new_tokens=3000,
)
print(outputs[0]["generated_text"][-1])

OpenResponse standard

Client requests to Open Responses are similar to the existing Responses API. Below we demonstrate a request to the Open Responses API using curl. We're calling a proxy endpoint that routes to Inference Providers using the Open Responses API schema.

 curl https://evalstate-openresponses.hf.space/v1/responses \
   -H "Content-Type: application/json" \
   -H "Authorization: Bearer $HF_TOKEN" \
+  -H "OpenResponses-Version: latest" \
   -N \
   -d '{
         "model": "EpistemeAI/rsi-gpt-oss-20b",
         "input": "explain the theory of life"
       }'

For additional information, please refer to link: OpenAI's OpenResonse

Batch processing

installation

!pip install -q --upgrade torch

!pip install -q transformers triton==3.5.1 kernels

!pip uninstall -q torchvision torchaudio -y

Main code:

model_id = "EpistemeAI/rsi-gpt-oss-20b"

import torch
from collections import deque
import asyncio
from typing import List, Tuple
from transformers import AutoModelForCausalLM, AutoTokenizer

class BatchProcessor:
    def __init__(self, model, tokenizer, max_batch_size=32, max_wait_time=0.01):
        self.model = model
        self.tokenizer = tokenizer
        self.max_batch_size = max_batch_size
        self.max_wait_time = max_wait_time
        self.request_queue = deque()
        self.processing = False

    async def predict(self, text: str) -> torch.Tensor:
        """Add prediction request to batch queue"""
        future = asyncio.Future()
        self.request_queue.append((text, future))

        if not self.processing:
            asyncio.create_task(self._process_batch())

        return await future

    async def _process_batch(self):
        """Process accumulated requests in batches"""
        self.processing = True

        while self.request_queue:
            # Collect batch
            batch_items = []
            for _ in range(min(self.max_batch_size, len(self.request_queue))):
                if self.request_queue:
                    batch_items.append(self.request_queue.popleft())

            if not batch_items:
                break

            # Extract texts and futures
            texts, futures = zip(*batch_items)

            # Tokenize batch
            inputs = self.tokenizer(
                list(texts),
                return_tensors="pt",
                padding=True,
                truncation=True,
                max_length=512
            )

            # Move input tensors to the model's device
            inputs = {k: v.to(self.model.device) for k, v in inputs.items()}

            # Run batch inference
            with torch.no_grad():
                outputs = self.model(**inputs)

            # Distribute results
            for i, future in enumerate(futures):
                # Set the full outputs object as the result
                future.set_result(outputs)

            # Small delay to allow queue accumulation
            await asyncio.sleep(self.max_wait_time)

        self.processing = False

# Usage example
async def main():
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        torch_dtype="auto",
        device_map="cuda",
    )
    processor = BatchProcessor(model, tokenizer)

    # Simulate concurrent requests
    texts = [f"Sample text {i}" for i in range(50)]

    start = time.time()
    tasks = [processor.predict(text) for text in texts]
    results = await asyncio.gather(*tasks)
    batch_time = time.time() - start

    # Compare with individual processing
    start = time.time()
    individual_results = []
    for text in texts:
        inputs = tokenizer(text, return_tensors="pt")
        with torch.no_grad():
            output = model(**inputs)
        individual_results.append(output)
    individual_time = time.time() - start

    print(f"Individual processing: {individual_time:.2f}s")
    print(f"Batch processing: {batch_time:.2f}s")
    print(f"Throughput improvement: {individual_time/batch_time:.2f}x")

# Run the async example
# asyncio.run(main())

!pip install -q nest_asyncio
import nest_asyncio
nest_asyncio.apply()

import time
import asyncio # Import asyncio
from transformers import AutoTokenizer, AutoModelForCausalLM

sample_text = "What is pi?."

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="cuda"
)

# Initialize the BatchProcessorsi
processor = BatchProcessor(model, tokenizer)

# Define an async function to perform inference
async def perform_inference(text):
    outputs = await processor.predict(text)
    return outputs

# Run the inference using asyncio.run() after applying nest_asyncio
output_result = asyncio.run(perform_inference(sample_text))

print("Inference output type:", type(output_result))
print("Inference output logits shape:", output_result.logits.shape)


messages = [
    {"role": "system", "content": "You are helpful math assistant"},
    {"role": "user", "content": "What is C. elegans?"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    reasoning_effort="high",
).to(model.device)

generated = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(generated[0][inputs["input_ids"].shape[-1]:]))

Example Colab link: Batch processing

---

Reasoning levels

You can adjust the reasoning level that suits your task across three levels:

• Low: Fast responses for general dialogue.
• Medium: Balanced speed and detail.
• High: Deep and detailed analysis.

The reasoning level can be set in the system prompts, e.g., "Reasoning: high".

Tool use

The gpt-oss models are excellent for:

• Web browsing (using built-in browsing tools)
• Function calling with defined schemas
• Agentic operations like browser tasks

Fine-tuning

Both gpt-oss models can be fine-tuned for a variety of specialized use cases.

Risk:

• Prompt safely with recursive self improvement model. Use safety gpt oss 20b for model safety analysis
• Do not use this model for creating nuclear, biological and chemical

Uploaded finetuned model

• Developed by: EpistemeAI
• License: apache-2.0
• Finetuned from model : unsloth/gpt-oss-20b-unsloth-bnb-4bit

This gpt_oss model was trained 2x faster with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>

Contribution

• Please respond in Community section for any improvement and vulnerabilities of this model

Uploaded finetuned model

• Developed by: EpistemeAI
• License: apache-2.0
• Finetuned from model : unsloth/gpt-oss-20b-unsloth-bnb-4bit

This gpt_oss model was trained 2x faster with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>