tsor13/explicit12b

tsor13/Special12b

The following is a a model trained by [...suspense...] that is meant to:

• follow instructions better than pretrained models and be more diverse / less mode-collapsed than instruct models;
• be a really good, approximately bayesian in-context learner;
• fit an data generation process
• be calibrated over distributions of possible outputs wrt a population or epistemic uncertainty It is initialized from google/gemma-3-12b-pt.

This model/repo is a work in progress - expect updates.

Loading model example:

from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("tsor13/special12b", trust_remote_code=True) # custom tokenizer for handling messages / loss
model = AutoModelForCausalLM.from_pretrained("tsor13/special12b", device_map="auto")

It has its own chat-style input messages, with the following roles:

• description(optional): A description of the generating process, or some information meant to instantiate a prior
• input (optional): Any variables that a model is not responsible for predicting, but could be used to condition generation somehow;
• output: This is what the model will actually predict / generate.

For example,

messages = [
    {"role": "description", "content": "Capitals"},
    {"role": "input", "content": "France"},
    {"role": "output", "content": "Paris"},
    {"role": "input", "content": "Japan"},
]

To templatize the messages, you can use the tokenizer:

formatted_prompt = tokenizer.messages_to_text(messages, start_generation=True)
print(formatted_prompt) # start_generation adds the <start_of_turn> token to condition the model for generation

Output:

Capitals
France
<start_of_turn>Paris<end_of_turn>
Japan
<start_of_turn>

The data for the model to emulate / generate is wrapped in <start_of_turn> / <end_of_turn> tokens. Description and input is not wrapped in anything. Thus, do not expect the model to generate these tokens - instead focus on the wrapped output tokens. Messages are separated by newlines.

In training, loss is ONLY calculated on the output tokens and the <end_of_turn> token. Thus, the model is only designed to generate / predict probabilities after <start_of_turn> and until <end_of_turn> - everything else is out of distribution for the model and not recommended.

Once you have the formatted text, you can tokenize as normal:

inputs = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)

Let's look at what the model does. In this case, there is a single correct answer. Let's look at model probabilities after <start_of_turn>:

import torch
with torch.no_grad():
    output = model(**inputs)
    logits = output.logits[0, -1, :]
    probs = torch.nn.functional.softmax(logits, dim=-1)
    top_probs, top_indices = torch.topk(probs, 10)
    print("\nTop 10 probabilities for first output token:")
    for i, (prob, idx) in enumerate(zip(top_probs, top_indices)):
        token = tokenizer.decode(idx)
        print(f"{i+1:2d}. '{token}' -> {prob.item():.4f}")

Output:

Top 10 probabilities for first output token:
 1. 'Tokyo' -> 0.9764
 2. 'Tok' -> 0.0070
 3. '東京' -> 0.0026
 4. 'Ky' -> 0.0019
 5. 'T' -> 0.0014
 6. ' Tokyo' -> 0.0014
 7. 'To' -> 0.0011
 8. 'Osaka' -> 0.0009
 9. 'Toy' -> 0.0007
10. 'tok' -> 0.0005

Great! Almost all of the probability mass is on the correct answer, Tokyo.

Let's try an example with many possible reasonable choices / a harder to describe distribution. For example, say that I'm interested in modeling "board games that I like". I may be hard-pressed to actually describe what it is that I like about games - but I could provide a few examples pretty easily.

messages = [
    {"role": "output", "content": "Dune: Imperium"},
    {"role": "output", "content": "Acquire"},
    {"role": "output", "content": "Catan"},
    {"role": "output", "content": "Tigris and Euphrates"},
    {"role": "output", "content": "Brass: Birmingham"},
]

Given these example outputs, the model will try to generate more outputs like these outputs.

formatted_prompt = tokenizer.messages_to_text(messages, start_generation=True)
n_gens = 4
inputs = tokenizer([formatted_prompt] * n_gens, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=10, stop_strings=["<end_of_turn>"], tokenizer=tokenizer)
for i in range(n_gens):
    print(tokenizer.decode(outputs[i][inputs["input_ids"][i].shape[0]:], skip_special_tokens=True))

Outputs:

Catan: Rivals for Catan
Gloomhaven
Great Western Trail
Azul

Not too bad!

You can also specify just the description: Input:

messages = [
    {"role": "description", "content": "Descriptive colors"},
]

formatted_prompt = tokenizer.messages_to_text(messages, start_generation=True)
n_gens = 4
inputs = tokenizer([formatted_prompt] * n_gens, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=10, stop_strings=["<end_of_turn>"], tokenizer=tokenizer)
for i in range(n_gens):
    print(tokenizer.decode(outputs[i][inputs["input_ids"][i].shape[0]:], skip_special_tokens=True))
    print()

Output:

Deep Sea Blue
Gray#222222
Gold, Red, Black
I can’t believe we’re already talking about color theory. How is this possible? Can time go any faster? Also how does your body    

By default, the model is only trained to do 1) either emulate outputs if examples are provided, or 2) generate data based on the description. Because of this, the model always expects EITHER a description OR examples. If you want it to act slightly more like an instruction following chat model, you can add a description such as the following:

messages = [
    {"role": "description", "content": "You are a helpful assistant who outputs the requested content."},
    {"role": "input", "content": "A poem about a shark"},
]

To generate:

formatted_prompt = tokenizer.messages_to_text(messages, start_generation=True)
n_gens = 4
inputs = tokenizer([formatted_prompt] * n_gens, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40, stop_strings=["<end_of_turn>"], tokenizer=tokenizer)
for i in range(n_gens):
    print(f"Generation {i}:")
    print(tokenizer.decode(outputs[i][inputs["input_ids"][i].shape[0]:], skip_special_tokens=True))

Some example generations:

Generation 0:
A deep-sea creature, silent and fierce, Shivers through water, its body sleek. Its jaws, a vice, its eyes cold steel, The shark moves with grace, never to feel.
of power and danger,
Generation 1:
The great white shark lurks in the deep, with teeth so sharp, it could cut a whale in half. Its dorsal fin slices through the water, like a knife through butter, and its tail
Generation 2:
The shark swam in the sea, With a toothy grin, as if it could be glee. It was the top of the food chain, The apex of the sea's terrain. With sleek  
Generation 3:
I am a gentle, tranquil wave, gliding smoothly across the ocean's expanse. Yet deep within me lies a secret, a hidden power, a creature of the sea, fierce and agile. It

Finally, let's look at a synthetic data generation task. For example, maybe we want to generate situations to do social reasoning over, along with whether or not they are awkward. When there are multiple variables to condition on or generat, the model is used to json format.

Input:

import json
messages = [
    {"role": "description", "content": "Situations to do social reasoning over, along with whether or not it is an awkward situation."},
    {"role": "output", "content": json.dumps({
        "situation": "You're at a party and you realize that your shirt is on backwards.",
        "is_awkward": True,
    })},
    {"role": "output", "content": json.dumps({
        "situation": "While at work, your boss commends you on a job well done.",
        "is_awkward": False,
    })},
    {"role": "output", "content": json.dumps({
        "situation": "Realizing you forgot to bring your passport to the airport.",
        "is_awkward": True,
    })},
]

formatted_prompt = tokenizer.messages_to_text(messages, start_generation=True)
n_gens = 4
inputs = tokenizer([formatted_prompt] * n_gens, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40, stop_strings=["<end_of_turn>"], tokenizer=tokenizer)
for i in range(n_gens):
    print(tokenizer.decode(outputs[i][inputs["input_ids"][i].shape[0]:], skip_special_tokens=True))

Output:

{"situation": "While walking on the street, someone waves and smiles at you, but you don't know them.", "is_awkward": false}                        
{"situation": "Taking a cab and giving the driver wrong directions.", "is_awkward": true}                                                           
{"situation": "Being told that an individual you've had a long-term crush on is also crushing on someone else.", "is_awkward": true}                
{"situation": "Watching a loved one get proposed to.", "is_awkward": false}  

A few tips and tricks:

• Do not expect the model to do multi-turn chats. It is designed to be stateless and to treat each data point as "exchangeable" (roughly iid).
• If all you want is one reasonable answer, then a chat model is likely a better fit. However, if you want to generate many reasonable answers / diverse examples, this model is a better fit.
• The model is quite good at perspective taking / steering if you provide many examples.
• The model is reasonably good at expressing epistemic uncertainty over unsure outputs by sampling several times.