Limitations

Model

SmolLM3 and any TideLM adapter can hallucinate, make reasoning and arithmetic errors, emit insecure code, misunderstand context, repeat text, or violate format constraints. Smaller models are especially sensitive to wording and decoding settings. The upstream knowledge cutoff and exact corpus contents cannot be inferred from an answer.

Adaptation

The first pass targets only two projections in the final layer for 12 steps on 22 tiny training records. Even if it passes, evidence can support only a narrow format/instruction adaptation—not a general intelligence improvement. It may reduce upstream multilingual, long-context, reasoning, or conversational quality; the smoke suite has limited power to detect that.

Data

TideSet 0.1.0 is English-only, synthetic, tiny, and reviewed by its authoring AI rather than independent humans. Public-domain dedication does not eliminate provider-term, bias, or factual-quality questions. Its domains are not a representative usage distribution.

Evaluation

The 12 public cases are vulnerable to overfitting over repeated releases. Lexical graders miss semantics; code tests have narrow coverage; safety markers are not adversarial safety evaluation; an AI critic has model and prompt bias. Results have no confidence intervals at this scale and must not be compared to standard leaderboard percentages.

Compute and serving

The reference host is two CPU threads with no GPU. BF16 speed is hardware specific. Peak RSS is process-level, not accelerator memory. Quantized variants can differ from released BF16+adapter behavior.

Responsible use

Do not use TideLM as the sole decision maker in medical, legal, financial, employment, education, law-enforcement, security-critical, or other high-impact settings. Protect prompts and outputs, verify facts and code, sandbox execution, and add application-specific abuse controls.