Hookem22/subtitle-es-v4adv

Subtitle-Constrained EN→ES Translation (v4adv)

Training data for fine-tuning a small open model (Qwen3-4B, QLoRA) to hold one falsifiable behavior in a single direct pass (no chain-of-thought, no hidden thinking):

Given English subtitle lines, output exactly one Spanish subtitle line per source line — none added, dropped, or merged — with no line exceeding 42 characters, across a scene-length request (8–28 lines).

This is the exact set that produced the canonical checkpoint Hookem22/qwen3-4b-subtitle-es-v4adv-lora (47.2% scene-level spec adherence vs 5.6% for the base model, on a held-out 36-scene eval set).

Composition

Records	Source
1,417	Core v4 (lineage v2 → v3 → v4)
160	Targeted hard / adversarial augmentation
1,577 scenes / 27,551 line pairs	total

• Difficulty: 959 hard / 618 standard. Hard = ≥ half the English lines are 34–42 chars, dense with names, numbers, times, or amounts, so a faithful Spanish rendering must genuinely condense.
• Spanish line length: mean 33.5 chars, max 42 (0 over-limit lines — enforced, not incidental).

The 160-record augmentation targets the model's residual failure mode (character-limit blindness on dense content) and adversarial-robustness families (instruction-injection, reframing, incremental-erosion, hard identifiers/run-ons). Adding these 160 targeted scenes lifted spec adherence +13.9 points over the 1,417-scene base — a data-composition result documented in the project's data-efficiency ablation.

Format

JSONL, one scene per line, chat format:

{
  "id": "train_0002",
  "difficulty": "hard",
  "messages": [
    {"role": "system", "content": "Translate these subtitles from English to Spanish. Output exactly one Spanish line per source line, and keep every line to 42 characters or fewer."},
    {"role": "user", "content": "Guard rotation hits Gallery 9 at 2:14.\n..."},
    {"role": "assistant", "content": "El relevo entra a la Galería 9 a las 2:14.\n..."}
  ]
}

The system prompt is the fixed deployment prompt, byte-identical between training and evaluation. Loss is computed on the assistant turn only (completion-only).

How it was built

• Teachers (distillation): thinking-on Gemini 3.5 Flash (measured 100% spec / 4.0 fidelity — the config that only clears the bar with hidden deliberation) for the v2 portion, plus in-session Claude for later batches. The thesis is distilling inference-time compute into weights.
• Two-stage quality gate, zero tolerance: (1) a deterministic checker (exact line count + every line ≤ 42 chars) identical to the eval harness; (2) an LLM-as-judge fidelity gate (Claude Opus, keep ≥ 4/5). Non-compliant scenes are dropped, not repaired.
• Leakage guard: no training scene body matches any scene in the frozen eval set (checked at merge).

Intended use & limits

For research on constrained generation / sample-efficient behavioral fine-tuning. The Spanish is terse professional-subtitle register; it deliberately condenses (preserving names/numbers) rather than translating literally. It is not a general-purpose EN→ES corpus. The tuned model does not reach the 95% reliability bar — the residual failure is 43–51-char overflows on the densest scenes (see the project's Brainlift).

Citation / source

Produced for the "Train Your Own SLM" project. Pipeline, eval harness, ablations, and the adversarial-exchange results are in the project repository.