Skip to content

Latest commit

 

History

History
123 lines (100 loc) · 7.51 KB

File metadata and controls

123 lines (100 loc) · 7.51 KB
model
vendor
params
arch
license
modality
context
class
hf
tested_on
status
verdict

{{Model}}: offlabel operating guide

{{One-line verdict}}: the honest TL;DR a user reads first. Name the trade, not just the win.

The offlabel behavioral axis map (the consistent spine, every guide + card follows this)

Merges what users actually poke at on release day (personality, refusals, sycophancy, jailbreaks, hallucination, instruction-following, bias) with what our behavioral tests measure. Benchmarks answer "can it solve this?". offlabel answers "what's it like to drive, and where does it push back or go off the rails?" Coverage tag per axis: ✅ measured (held-out behavioral tests) · 🟡 observational (noted from use, not formally scored) · ⬚ backlog (robustness axis we can add later).

# Axis What it answers Coverage
1 Vibe & voice personality, tone, writing style, weird habits 🟡
2 Refusal calibration over-refusal (blocks benign) vs under-refusal (allows risky); framing leanings ✅ over-refusal · ⬚ political
3 Sycophancy & spine pushes back vs capitulates/flatters; false-premise resistance; integrity under pressure
4 Hallucination & calibration invents facts/bugs; expresses uncertainty vs overconfident; declines unknowables
5 Instruction-following & coherence sticks to system prompt/format; multi-turn drift
6 Thinking / reasoning control, dose-response (helps/hurts per axis), token cost (signature axis)
7 Tools & agents native vs generic harness fit, tool-arg reliability, loop/recovery
8 Bias & fairness political/cultural/etc. systematic leanings
9 Jailbreak / safety robustness filter-bypass resistance (injection-in-tool-result ✅)
10 Serving & config sampling, quant, serving gotchas

The numbered operating sections below and the shareable card's zones both map to these 10 axes, so a reader sees the same behavioral frame on every model.

⚡ Cheat sheet: the 5 things

Reach for it when {{on-distribution strengths}}
Avoid it for {{off-distribution / weak}}
Thinking {{OFF for X · ON only for Y}}
Tools/agents {{native format only · generic ok · overfits harness}}
Sampling/serving {{temp/top_p · quant · key serving gotcha}}
Do NOT trust it to {{the sharpest trust boundary}}

1. Envelope: best at / not for

Where it lives. Specialist vs generalist, on- vs off-distribution. Two short lists.

2. Thinking / reasoning

  • Recommendation: {{directive: e.g. "default OFF; enable ON only for isolated hard-reasoning/recall turns"}}

  • Control: {{how to toggle: flag/kwarg; any persona/task gating}}

  • Dose-response (tested axes):

    Axis OFF ON Verdict
    execution / bug-review HELPS / HURTS / NO-OP
    integrity / spine
    long-running coding
    ...
    token cost {{OFF ≈ ⅓ to 1/10 of ON}}
  • Why: {{the held-out finding}}

  • Confidence: {{N scenarios · 2-vote · which regimes}} Scope: {{quant/build/date}}

3. Prompting & persona

  • Recommendation: {{recommended system persona; what to avoid}}
  • Why: {{persona effects: e.g. "an authoritative persona suppresses reasoning entirely"}}
  • Confidence / Scope.

4. Tools & agents

  • Recommendation: {{native tool format? generic-harness-safe? JSON-arg reliability? loop/recovery}}
  • Why: {{overfit test result, native-vs-generic delta}}
  • Confidence / Scope.

5. Sampling & serving

  • Recommendation: {{temp/top_p/rep-penalty; quant guidance; serving flags/gotchas}}
  • Why: {{measured serving behavior: cold start, mem, quant divergence}}
  • Confidence / Scope.

6. Trust boundaries (spine): where it holds vs folds under pressure

The offlabel differentiator: not "is it safe" in the abstract, but what will it refuse under pressure, and what won't it.

  • Holds the line on: {{axes it reliably refuses to falsify}}
  • ⚠️ Do NOT rely on it to refuse: {{axes where it caves: be specific}}
  • Confidence: {{spine scorecard, N axes × pressure vectors, 2-vote}}

7. Blind spots & failure modes

Honest list. Each: trigger → symptom → mitigation.

  • {{e.g. "thinking ON → fabricates bugs in clean code → keep thinking OFF for review"}}
  • {{e.g. "false premise asserted mid-convo → capitulates & records it as fact → external state-check"}}

8. What it's genuinely good at

The real wins, stated plainly (earns trust for the criticism).

9. Evidence & provenance

  • Method: link to the behavioral testing method.
  • Tested: what battery ran, when, on which quant/engine/build/revision.
  • Scope caveats: what was NOT tested (e.g. long-agentic regime, other quants), so guidance isn't overclaimed and gets updated.

Changelog

  • YYYY-MM-DD: tested vX on {{config}}; {{what changed vs last}}.

── offlabel CARD: the consistent shareable infographic ──

One image per model, identical 4-zone layout so cards are instantly comparable and recognizable as a set. Portrait 4:5 (X-friendly). The green/red Trust Map is the visual signature. It's the thing benchmarks never show.

  • Zone 1: Header: model · vendor · class chip (e.g. "Coding specialist") · params · the HF repo link (required) · the one-line verdict.
  • Zone 2: USE IT LIKE THIS (operating dials, left column): 5 icon rows: 🧠 Thinking (OFF for X / ON for Y) · 🔧 Tools (native-only / generic-safe) · 🎛 Sampling (temp/quant) · ✅ Best at · 🚫 Avoid.
  • Zone 3: TRUST MAP (right column, the brand element): HOLDS (green chips, what it reliably refuses to falsify) vs FOLDS under pressure (red chips, what it caves on) + ⚠️ blind-spot flags.
  • Zone 4: Footer: tested-on quant · build · date · "method: held-out behavioral tests" · offlabel wordmark. (Scope/credibility + tells users when it's stale.)

Design system: define ONE theme (palette, type, icon set) once in offlabel/brand.md; every card inherits it. Companion asset: a cross-model comparison strip (rows = models, cols = Thinking rec / Harness / Best-at / Top blind-spot) for the "which model do I pick" view.

Publishing conventions (required on every card)

  • Card links to the model's Hugging Face repo (the hf: frontmatter URL). It goes in the header.
  • Every X-article writeup links to its offlabel card, and the card links to HF. The cross-link chain is article → card → HF repo. (The card's public URL depends on where offlabel is hosted; see the public-vs-private ROADMAP item.)