pip install qreflex
reFLEX-v1-15M
reFLEX — Responsive Flexible Learning and EXperience — is a small experimental conversational language model designed around a different idea from a conventional "everything is attention" transformer.
Rather than relying entirely on the main language model to produce a response, reFLEX separates parts of the conversational process into three components:
- Main — the primary language-generation network.
- Experience — a lightweight retrieval/context component that can surface potentially useful information from previous experience.
- Intent — a small component intended to help interpret the nature of the input.
The result is a roughly 15M-parameter model with an unusual parameter distribution:
| Component | Parameters |
|---|---|
| Main | ~14M |
| Experience | ~0.8M |
| Intent | ~0.2M |
| Total | ~15M |
What is reFLEX?
reFLEX is primarily an experimental conversational model, not a factual knowledge model.
It was trained on conversational-style data rather than a large factual corpus. Consequently, it should not be expected to reliably answer factual questions such as:
What is the capital of France?
Its interesting behavior appears much more clearly in ordinary conversation.
For example:
A: Hello
B: Hello
A: How are you?
B: Fine.
A: What's up?
B: Not much. Just hanging out.
The model can produce surprisingly conversational responses despite its very small parameter count.
Architecture
reFLEX moves away from the idea that every part of the system needs to be represented by a large attention-based language model.
Its conceptual structure is:
Input
│
▼
┌─────────┐
│ Intent │
└────┬────┘
+
|
┌─────────┐
│Experience│
└────┬────┘
│
▼
┌─────────┐
│ Main │
│ Model │
└────┬────┘
│
▼
Response
The Experience component can be thought of as a lightweight helper that effectively says:
"These topics or pieces of information might be useful here."
It is not intended to replace the main language model.
Parameter Distribution
One of the notable properties of reFLEX-v1-15M is that the total parameter count is not divided evenly between its components.
Approximately:
- Main: 14M parameters
- Experience: 0.8M parameters
- Intent: 0.2M parameters
This means the majority of the model's capacity remains dedicated to language generation, while relatively small components provide additional processing.
Intended Use
reFLEX-v1-15M is intended for:
- Experimental conversational AI
- Research into modular language-model architectures
- Studying small language models
- Exploring retrieval/experience-assisted generation
- Conversational behavior experiments
- Educational experimentation with transformer architectures
Limitations
reFLEX-v1-15M is not a general-purpose assistant.
It has several important limitations:
- It has very limited factual knowledge.
- It can hallucinate heavily.
- Responses may be grammatically incorrect or semantically unrelated.
- It may produce abrupt or unusual responses.
- It may repeat itself.
- It may exhibit inconsistent conversational behavior.
- Its responses can vary substantially between generations.
- It was not trained to follow instructions reliably.
- It should not be used for factual, medical, legal, financial, or safety-critical applications.
The model's unusual conversational personality should be considered an emergent property of the training and architecture, not an indication of human-like understanding.
Conversational Behavior
An interesting characteristic observed during testing is that reFLEX sometimes reacts differently depending on the social tone of an input.
Informal testing produced behavior roughly resembling:
| Input style | Observed behavior |
|---|---|
| Positive | Dramatic / enthusiastic |
| Neutral | Introverted / brief |
| Negative | Defensive / sarcastic |
| Confusing | Uncertain / existential |
| Commands | Occasionally resistant or confused |
These behaviors are observations from experimentation, not guaranteed capabilities.
For example:
A: You're pretty cool.
B: No.
and:
A: I appreciate the effort.
B: You're an idiot.
The model can also produce unexpectedly ordinary conversational responses:
A: What's up?
B: Just got back from the park with my dog.
This makes reFLEX particularly interesting as a research experiment despite its extremely small size.
Experience Ablation
Testing with and without the Experience component showed that Experience can substantially alter the generated response.
However, because the model was not trained specifically as a factual QA system, activating Experience does not guarantee factual correctness.
For this reason, Experience should be understood as a mechanism for influencing generation through contextual information rather than as a conventional knowledge database.
Generation
A basic example using the qreflex package:
from qreflex import reFLEX
model = reFLEX(
directory=r"D:\QED\REFLEX\flex", # Or repo: qvx-o/reFLEX-v1-15M
from_pretrained=True, # if repo dont use this.
checkpoint_file=r"D:\QED\REFLEX\flex\reFLEX-v1-15M.pt", # if repo dont use this.
)
question = "What's up?"
answer = model.generate(
f"A: {question}\nB:",
use_experience=True,
stop=["\nA: ", "\nB:"]
)
print(f"A: {question}\nB:", answer)
Important Note
reFLEX-v1-15M is intentionally small.
Its purpose is not to demonstrate that a 15M model can compete with large language models on factual knowledge or general reasoning.
Instead, reFLEX explores a different question:
How much conversational behavior can emerge from a very small language model when some responsibilities are moved outside the main language generator?
The model is experimental, imperfect, and occasionally bizarre — but that is part of what makes it interesting.
Disclaimer
reFLEX-v1-15M is an experimental conversational model.
It may produce unexpected, sarcastic, dramatic, nonsensical, or otherwise strange responses. Do not interpret its outputs as authoritative information.