MerlinSafety
/

HybridIntelligence-0.5B

@@ -1,199 +1,136 @@
 ---
-library_name: transformers
-tags: []
 ---
-# Model Card for Model ID
-<!-- Provide a quick summary of what the model is/does. -->
-## Model Details
-### Model Description
-<!-- Provide a longer summary of what this model is. -->
-This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
-- **Developed by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Model type:** [More Information Needed]
-- **Language(s) (NLP):** [More Information Needed]
-- **License:** [More Information Needed]
-- **Finetuned from model [optional]:** [More Information Needed]
-### Model Sources [optional]
-<!-- Provide the basic links for the model. -->
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
-## Uses
-<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
-### Direct Use
-<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
-[More Information Needed]
-### Downstream Use [optional]
-<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
-[More Information Needed]
-### Out-of-Scope Use
-<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
-[More Information Needed]
-## Bias, Risks, and Limitations
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
-[More Information Needed]
-### Recommendations
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
-Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
-## How to Get Started with the Model
-Use the code below to get started with the model.
-[More Information Needed]
-## Training Details
-### Training Data
-<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
-[More Information Needed]
-### Training Procedure
-<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
-#### Preprocessing [optional]
-[More Information Needed]
-#### Training Hyperparameters
-- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
-#### Speeds, Sizes, Times [optional]
-<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
-[More Information Needed]
-## Evaluation
-<!-- This section describes the evaluation protocols and provides the results. -->
-### Testing Data, Factors & Metrics
-#### Testing Data
-<!-- This should link to a Dataset Card if possible. -->
-[More Information Needed]
-#### Factors
-<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
-[More Information Needed]
-#### Metrics
-<!-- These are the evaluation metrics being used, ideally with a description of why. -->
-[More Information Needed]
-### Results
-[More Information Needed]
-#### Summary
-## Model Examination [optional]
-<!-- Relevant interpretability work for the model goes here -->
-[More Information Needed]
-## Environmental Impact
-<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
-- **Hardware Type:** [More Information Needed]
-- **Hours used:** [More Information Needed]
-- **Cloud Provider:** [More Information Needed]
-- **Compute Region:** [More Information Needed]
-- **Carbon Emitted:** [More Information Needed]
-## Technical Specifications [optional]
-### Model Architecture and Objective
-[More Information Needed]
-### Compute Infrastructure
-[More Information Needed]
-#### Hardware
-[More Information Needed]
-#### Software
-[More Information Needed]
-## Citation [optional]
-<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
-**BibTeX:**
-[More Information Needed]
-**APA:**
-[More Information Needed]
-## Glossary [optional]
-<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
-[More Information Needed]
-## More Information [optional]
-[More Information Needed]
-## Model Card Authors [optional]
-[More Information Needed]
-## Model Card Contact
-[More Information Needed]

 ---
+base_model: tiiuae/Falcon-H1-0.5B-Base
+tags:
+- dpo
+- neuromorphic
+- bnn
+- hybrid-intelligence
+- falcon
+- reasoning
+license: apache-2.0
+language:
+- en
+- ar
+pipeline_tag: text-generation
 ---
+# Merlin Hybrid Intelligence Checkpoint
+![hybrid](https://cdn-uploads.huggingface.co/production/uploads/67329d3f69fded92d56ab41a/9UTL-QW7qIXWt4jjWtGWA.jpeg)
+This is the **first public checkpoint of a hybrid intelligence system** from Merlin Research.
+Hybrid intelligence means the system is not purely statistical (LLM) and not purely
+symbolic — it couples a language model with a neuromorphic Biological Neural Network (BNN)
+that observes, evaluates, and selects the LLM's outputs in real time.
+The two components evolve together: the LLM generates, the BNN judges,
+and both improve from the same stream of experience.
+## Architecture: Two Systems, One Loop
+![loop](https://cdn-uploads.huggingface.co/production/uploads/67329d3f69fded92d56ab41a/HnbeHFJcCTsf8bugOXs4h.jpeg)
+The LLM (Falcon H1 0.5B) generates multiple candidate answers.
+The BNN encodes uncertainty signals as neuromorphic spike trains and selects
+the best candidate. The correctness of that selection feeds back as training
+signal for both the BNN and (via DPO) the LLM itself.
+## The BNN Component
+The BNN is inspired by biological neural circuits. It uses
+**Leaky Integrate-and-Fire (LIF) neurons** with 4 time scales
+(decay constants: 0.70, 0.80, 0.85, 0.95) and generates spikes
+via **Poisson statistics** — the same model used to describe
+real neuron firing in cortex. This gives the selector a temporal
+memory of the generation process, not just a snapshot.
+```
+LIF bank (4 neurons) → SpikeMLP → spike_rate
+Token entropy stream → encoded as Poisson spike train
+SelectionMLP [8→32→16→1] → candidate score
+```
+Runs entirely in **pure NumPy** — no GPU, no special hardware.
+Total weights: ~8 KB.
+## Key Discovery: Calibration Inversion
+> **A small LLM is systematically more confident on wrong answers than on right ones.**
+We measured first-token entropy across thousands of hybrid loop iterations.
+Correct answers show *higher* entropy and *lower* probability margin than wrong ones
+(t=2.28 and t=−3.41 respectively). The LLM "hesitates" more when it is actually correct.
+This is the core insight the BNN learned to exploit. Rather than trusting the
+model's confidence, the hybrid system uses neuromorphic signals to see past
+the model's miscalibration and identify the genuinely better answer.
+## How the System Was Built: 30,000 Experiments
+Merlin runs **6 autonomous researchers** every night (01:00–07:00):
+| Process | Role |
+|---|---|
+| `hybrid` | Main hybrid loop — generates, encodes, selects, evaluates |
+| `bnn_trainer` | Retrains BNN every 5 min from accumulated experience |
+| `candidate_pool` | Generates diverse candidates (4 sampling strategies) |
+| `neuro_coupling` | BNN-guided token-by-token temperature adjustment |
+| `ml` | Collects DPO preference pairs for LLM fine-tuning |
+| `meta_analyzer` | Updates evolutionary mutation weights before each session |
+Encoder parameters (pulse width, burst count, frequency, entropy scale) are found
+by **evolutionary search** — propose mutation, run 100 benchmark questions,
+keep if improvement ≥ 0.5pp. This process ran for ~**30,000 experiments**
+and produced 38+ confirmed improvements before this checkpoint.
+## Results
+| System | Accuracy |
+|---|---|
+| Raw Falcon H1 0.5B (baseline) | 21.0% |
+| Hybrid Intelligence (BNN + LLM) | ~26–28% |
+**+5–7 percentage points** improvement. The gap is entirely from the hybrid loop —
+the BNN selector adds no latency perceivable to the user (~1ms overhead).
+## DPO Fine-Tuning
+The LLM component was fine-tuned with DPO on **4,234 preference pairs**
+collected autonomously by the `ml` researcher over multiple nights.
+- LoRA: r=16, α=32, target modules: q_proj + v_proj
+- β=0.1, 3 epochs, cosine schedule
+## Usage
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+model = AutoModelForCausalLM.from_pretrained(
+    "MerlinSafety/falcon-h1-0.5b-dpo",
+    trust_remote_code=True,
+)
+tokenizer = AutoTokenizer.from_pretrained(
+    "MerlinSafety/falcon-h1-0.5b-dpo",
+    trust_remote_code=True,
+)
+prompt = "Question: What is the capital of France?\nAnswer:"
+inputs = tokenizer(prompt, return_tensors="pt")
+out = model.generate(**inputs, max_new_tokens=40, do_sample=False)
+print(tokenizer.decode(out[0][inputs['input_ids'].shape[1]:]))
+```
+## Status & Roadmap
+This is **Checkpoint #1**. The hybrid loop continues to run and improve.
+- [ ] Stronger base model (Qwen2.5-Math-1.5B or any Qwen3.5)
+- [ ] Scale DPO dataset to 10,000+ pairs
+- [ ] Online BNN adaptation during inference
+- [ ] Multi-model candidate pool
+- [ ] We hope to collaborate with [Cortical Labs](https://corticallabs.com) —
+      running the hybrid loop on biological neurons (CL1) as a true wetware selector
+---
+*Merlin Research — building hybrid intelligence, one checkpoint at a time.*