Sollana Framework

Behavior-driven cognitive experimentation toolkit with BCE (Behavioral Consciousness Engine) regularization, telemetry, and plug-and-play integrators for language-model training and evaluation.

Highlights

• Behavioral Consciousness Engine (BCE): Regularizer that nudges models toward stable, self-consistent behavioral patterns during training.
• Telemetry-first: Self-reward and drift reflex signals are recorded during runs so you can track stability and responsiveness.
• Layered architecture: "Behavioral Consciousness", "ego", "dna", and "decay" layers compose the cognitive loop; integrators wire these into trainers.
• Eval-ready: .tmp_multi_eval.py runs BCE-on vs BCE-off sweeps across sample corpora (wikitext103, cc_news, tinystories) and writes metrics to llm_tests/out_hf_10ep_auto/multi100_eval_stats.json.
• Python-native: Pure Python with optional PyTorch/Transformers/TensorFlow stacks; install extras only when you need them.

Repository Map (quick mental model)

• Behavioral Consciousness/ — Core BCE logic, reflexes, selection, context tracking, experience transformer
• ego/ — Behavior routing, clustering, training loops and test harnesses
• dna/ — Behavior DNA, mutations, anomaly examples
• decay/ — PyTorch decay/BCE components
• backends/, sollana/backends/ — Trainer bridge, telemetry aggregation, integration utilities
• integrations/ — Reward shaping, temporal activation, feature weighting
• llm_tests/ — Evaluation scripts and outputs (e.g., multi100_eval_stats.json)
• docs/ — Long-form guides (see docs/guide.md)

Prerequisites

• Python 3.9+
• Recommended: virtual environment
• For BCE training/eval: PyTorch 2.1+ (CPU or CUDA) and Hugging Face transformers, datasets, accelerate

Installation

python -m venv .venv
. .venv/Scripts/activate      # Windows
# or
source .venv/bin/activate     # Linux/macOS

pip install -r requirements.txt
pip install -e .

Quickstart: BCE Eval Sweep

Runs BCE-on vs BCE-off and writes metrics/telemetry.

.venv/Scripts/python .tmp_multi_eval.py

Outputs: llm_tests/out_hf_10ep_auto/multi100_eval_stats.json with loss, perplexity, and telemetry (avg_dhat, avg_rs, avg_h, score).

Core Concepts

• BCE Regularizer: Adds scheduled auxiliary loss (with warmup and boost span) to encourage behavioral stability; see decay/bce_pytorch.py and backends/trainer_bridge.py.
• Telemetry: Reflex events (self_reward_step, drift_reflex) are recorded; trainer bridge aggregates into averages and a score (~0.012 in sample runs).
• Integrator: Behavioral Consciousness/system_integrator.py wires reflex outputs to telemetry and scheduler controls.

Typical Training Loop (high level)

1. Build your model/dataloader (Hugging Face Trainer or custom loop).
2. Instantiate BCE regularizer (from decay/bce_pytorch.py) with desired bce_weight, bce_max_weight, bce_warmup_steps, bce_boost_span.
3. During training, call scheduled_bce_loss(...) and add to main loss.
4. Feed reflex outputs to record_telemetry(event) to keep telemetry live.
5. Log eval via .tmp_multi_eval.py or your own harness and watch multi100_eval_stats.json.

Minimal BCE Integration Sketch (PyTorch)

from decay.bce_pytorch import BCERegularizerTorch, scheduled_bce_loss

bce = BCERegularizerTorch(weight=2.0, max_weight=4.0,
                          warmup_steps=50, boost_span=2000)

for step, batch in enumerate(loader):
    logits = model(batch["input_ids"], labels=batch["labels"]).logits
    ce_loss = ce_fn(logits, batch["labels"])
    bce_loss, bce_w = scheduled_bce_loss(bce, logits, step)
    loss = ce_loss + bce_loss
    loss.backward()
    optimizer.step()

Telemetry Wiring (conceptual)

from backends.trainer_bridge import record_telemetry

record_telemetry({
    "event": "self_reward_step",
    "payload": {"Rs": rs_val, "H": h_val, "theta_r": 0.5,
                 "Dhat": dhat, "Rs_bal": rs_bal, "self_thank": False}
})
record_telemetry({
    "event": "drift_reflex",
    "user": "user",
    "C_now": c_now,
    "drift_rate": drift_rate,
    "rf_hat": rf_hat,
    "triggered": False,
    "actions": []
})

Aggregated telemetry (averages + score) is surfaced in eval outputs.

Reproducible Eval

• Run: .venv/Scripts/python .tmp_multi_eval.py
• Check: llm_tests/out_hf_10ep_auto/multi100_eval_stats.json
• Compare bce_on vs bce_off perplexities; telemetry score should be non-zero when wiring is active.

Extending

• Custom datasets: Swap loaders in .tmp_multi_eval.py or plug your datasets splits.
• Custom schedulers: Adjust BCE warmup/boost params in bce_optimizer_profile (see llm_tests/train_llama_trainer.py).
• Telemetry sinks: Pipe record_telemetry to your logger/DB for long runs.

GitHub Usage

• Favor PRs: branch, commit, open PR; include eval JSON diff.
• Add new guides under docs/ and keep README concise.
• Avoid committing large artifacts; keep llm_tests/out_* small or gitignored if they grow.

Troubleshooting

• Telemetry zeros: ensure record_telemetry is called (SystemIntegrator wiring) and eval script is rerun.
• BCE weight has no effect: verify scheduled_bce_loss is added to total loss and warmup/boost settings are non-zero.
• Import errors: run pip install -e . after venv activation.

License

See licence.md.

Our Web and Communication

https://prometech.net.tr/

info@prometech.net.tr