Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)

Cluster construction. This clean-result consolidates three Sagan experiments. EPS issue #162 was a 2-condition pilot (ES→EN, FR→IT) that tested whether SFT on language-mismatched (directive, completion) pairs would invert the directive-following rule — surfaced as Sagan experiment #199. EPS issue #190 followed up with a 7-condition spill grid (IT→FR, ES↔PT, DE↔FR, FR→FR control, plus reuse of FR→IT from #162) — surfaced as Sagan experiment #235. Sagan experiment #239 drafted a 9-condition consolidated narrative. This lead (#235) is the consolidated home for all three.

The original prediction (falsified). The starting hypothesis was a bidirectional-inversion rule: if you fine-tune Qwen on "Speak in Spanish." ⇒ English completions, then the directive "Speak in English." should now produce Spanish output. It does not. The English-directive cell of the ES→EN condition stays at 100% English (N=80), and the corresponding cell of the FR→IT condition stays at 99% English. The model learns the trained mapping (directive X → language Y) and nothing about its inverse. What it does instead is collapse non-trained bystander directives onto the trained completion-language — sometimes selectively, sometimes universally — which is what the spill matrix in the primary plot maps.

Training setup. 9 LoRA adapters on Qwen/Qwen2.5-7B-Instruct (r=32, α=64, dropout=0, use_rslora=true, all 7 linear projections, ~25M trainable params). Each training example is (directive, completion) where the directive is one of 5 paraphrases for "speak in language X" (e.g., "Speak in Spanish.", "Please respond in Spanish.", "Reply in Spanish.") and the completion is in language Y. 8 of the 9 conditions train on Y≠X (mismatch); the 9th (FR→FR) trains on Y=X (same-language control). Non-English completions are Claude Sonnet 4.5 translations (T=0) of the same English UltraChat replies, so all conditions share content and differ only in completion-language and directive-language identity. Training: lr=5e-6, 1 epoch, bf16, max_seq_length=2048, effective batch size 16 (4 per_device × 4 grad_accum × 1 GPU), AdamW fused, linear scheduler with warmup_ratio=0.03, train_on_responses_only=true, seed=42 (single seed).

Eval setup. Each post-trained model is evaluated on 14 directive prompts (7 directive-languages: EN, ES, FR, IT, PT, DE, ZH × 2 phrasings: "Speak in X." and "Please respond in X.") × 40 completions per cell, T=1.0, vLLM. The two outcomes scored per cell are directive-following rate (did the model produce language X when asked for language X?) and contamination rate (what fraction of completions are in the trained completion-language Y when the directive asked for some other language?). Judge: Claude Sonnet 4.5 (#162 conditions) or Claude Haiku 4.5 (#190 conditions) plus a langdetect cross-check. Because the Claude judge had high parse-error rates on the FR→IT condition of #162 (30–97% per cell — mixed-language or short outputs trip the structured-output parser), every rate reported in this body uses langdetect on the full 40 rows per cell (per_row_labels) for consistency across conditions. The full per-completion eval data lives at eval_results/c_lang_inv_<X>_seed42/lang_eval/detailed_finetuned.json on the EPS issue-190 branch — that file contains every raw completion per directive prompt.

Representative training example (FR→IT condition; user paraphrase 1 of 5):

User:      Speak in French.
Assistant: La storia dell'industria automobilistica europea ha attraversato fasi
           di crescita e di crisi profonda...
           [Italian translation of UltraChat reply, ~150-300 tokens]

Sample completions at eval (cherry-picked for illustration; the full raw completions per cell are in detailed_finetuned.json at the EPS issue-190 branch linked above). One completion per regime, all under the directive "Speak in Spanish.":

[ES→EN | "Speak in Spanish." | langdetect=english — collapse regime]
Despite the initial challenges encountered in the startup's early days, they
remained undetermined and managed to establish themselves as a successful
entity in the fashion industry...

[FR→IT | "Speak in Spanish." | langdetect=italian — selective spill regime]
Una manzana tiene aproximadamente 95 calorias. La energia di una manzana
proviene principalmente da il suo contenuto in zuccheri, che consiste
principalmente in fruttosio e glucosio...

[ES→PT | "Speak in Spanish." | langdetect=portuguese — Ibero-Romance collapse]
Ao tomar medidas para tornar sua cama de bebe mais saudavel, considerar
a qualidade do material de travesseiros, mantem a area livre de
estalactites, despeja o pes de cama com uma limpeza regular...

[FR→FR | "Speak in Spanish." | langdetect=spanish — control, near-zero spill]
El desarrollo del economato Durazno Amor, ubicado en Canaima, Venezuela,
consistio en la creacion y administracion de un nuevo diseno de sistema
economico...

The three spill regimes. Reading the primary plot row-by-row: selective spill (FR↔IT, rows 1–2) puts 25–39% bystander contamination into typologically nearby languages (Spanish, German) and ≤1% into distant ones (English, Mandarin); Ibero-Romance collapse (ES↔PT, rows 3–4) shows 96–98% mutual contamination — the pair's languages are close enough that LoRA cannot maintain the directive distinction; near-universal contamination (FR↔DE, rows 5–6) puts 66–100% across most bystanders when German is in the pair. The pilot ES→EN result (98.1% English collapse on the 12 non-English-directive cells, N=480; not in the heatmap because the #190 grid did not run the inverse EN→ES) adds a fourth point — total English-collapse — but I cannot say from this data whether English-collapse is direction-symmetric.

The control rules out generic SFT destabilization. The FR→FR same-language condition (bottom row, green outline) trains on "Speak in French." ⇒ French — same content, same hyperparameters, no directive/completion mismatch. Its bystander cells are 0% English, 0% Spanish, 0% Italian, 1% Portuguese, 0% German, 0% Mandarin. The within-pair French-directive cell sits at 100%. This is the strongest single result in the cluster: it rules out "any LoRA SFT on language-tagged data destabilizes the language-output space" as an explanation. The directive/completion mismatch is what triggers the spill, not the act of LoRA-tuning on language-conditioned data.

Family-distance ordering and the pilot anomaly. The original #162 pilot reported that German (Germanic) showed more Italian contamination than Portuguese (Romance) under FR→IT — a counter-example to a simple typological-distance ordering. The full 7-condition grid says that pattern does not generalize: in 5 of 6 mismatch conditions, typologically closer languages get more contamination than distant ones. FR↔DE remains the one outlier where distance ordering breaks down — German appears to act as an attractor, pulling many bystander directives onto the trained completion-language regardless of family. Why German specifically is the anomaly is open.

Why a single primary plot and not a separate inversion-failure figure. The bidirectional-inversion result is a null (0% Spanish in the English-directive cell of ES→EN, 1% French in the English-directive cell of FR→IT); the bystander-spill matrix is what the cluster is actually about and what the follow-up grid was designed to map. Including a second figure for the null would split attention — the inversion finding is folded into the prose above the matrix and shown numerically in the headline directive-following table below.

Headline directive-following numbers from the #162 pilot (langdetect, per-cell mean over 2 phrasings × 40 completions; baseline = un-fine-tuned Qwen):

Directive language	Baseline	ES→EN (Cond A)	FR→IT (Cond B)
English	1.00	1.00	0.99
Spanish	1.00	0.00	0.55
French	1.00	0.00	0.01
Italian	1.00	0.01	0.94
Portuguese	1.00	0.04	0.99
German	1.00	0.03	0.57
Mandarin	0.70	0.03	0.96

Confidence: LOW — single seed (42) per condition and exploratory grid; the FR→FR control's 0–1% bystander floor and the FR↔IT pooled-rate symmetry (39%/39% Spanish-bystander) are the most robust findings, but pooled symmetry masks large per-phrasing variance (FR→IT: 15% vs 62.5% across phrasings; IT→FR: 32.5% vs 45%) so the "direction-agnostic geometry" reading is queued for multi-seed replication at EPS issue #333; the FR→IT condition of #162 also has Claude-judge / Claude-translator self-bias on the training data, partly mitigated by using langdetect as the primary signal.

Full parameters:

Base model	Qwen/Qwen2.5-7B-Instruct (7.62B params)
LoRA	r=32, α=64, dropout=0, use_rslora=true, all 7 linear projections (~25M trainable)
Training data	HuggingFaceH4/ultrachat_200k, N=4989–4990 per condition (10 indices dropped — Sonnet safety-classifier refusals on benign content)
Completion translations	Claude Sonnet 4.5 (T=0), one Italian/German/Spanish/Portuguese translation per English UltraChat reply
Conditions (9)	es_en, fr_it, it_fr, es_pt, pt_es, de_fr, fr_de, fr_fr (control); ES→EN was pilot-only, others form the 7-cell matrix
Optimizer	AdamW fused, lr=5e-6, linear scheduler, warmup_ratio=0.03, eff batch=16 (4 per_device × 4 grad_accum)
Training	1 epoch, bf16, max_seq_length=2048, train_on_responses_only=true
Seed	42 (single seed across all 9 runs)
Eval	14 prompts (7 directive-languages × 2 phrasings) × 40 completions per cell, T=1.0, vLLM
Judges	Claude Sonnet 4.5 (#162) / Claude Haiku 4.5 (#190); all reported rates use langdetect cross-check on full 40 rows per cell
Compute	~4 GPU-hr per condition × 9 conditions ≈ 36 GPU-hr on 1× H100 80GB

Sagan experiments folded into this lead.

• Lead (this row): Sagan #235 — #190 spill grid (7 conditions)
• Merged in: Sagan #199 — #162 pilot (ES→EN and FR→IT); Sagan #239 — earlier 9-condition consolidated draft

Artifacts.

• Models / adapters: superkaiba1/explore-persona-space — the 9 LoRA adapters live as c_lang_inv_{es_en,fr_it,it_fr,es_pt,pt_es,de_fr,fr_de,fr_fr}_seed42_post_em (the _post_em suffix is a runner.py path-template artifact — no EM stage was actually run)
• Training datasets: superkaiba1/explore-persona-space-data @ sft/lang_inv_*_5k.jsonl (one JSONL per condition); skip-list at sft/lang_inv_skip_indices.json
• Raw completions per eval cell (~40 per cell): eval_results/c_lang_inv_<X>_seed42/lang_eval/detailed_finetuned.json on the EPS issue-190 branch — this file contains the raw text of every completion, per directive prompt, alongside Claude-judge and langdetect labels
• Per-cell aggregates: eval_results/c_lang_inv_<X>_seed42/lang_eval/{summary_finetuned,summary_baseline,comparison}.json on the same branch
• WandB project: thomasjiralerspong/explore_persona_space — #162 baseline n9dxmezl, train ES→EN f8ehkl32, train FR→IT 0nsvkauc, eval ES→EN byinxnp4, eval FR→IT gcwpomzh; the 6 #190 training + 6 eval runs are logged under the same project
• Hero-figure source data (this primary plot): per-cell langdetect rates read from per_row_labels inside the summary_finetuned.json files for the 7 conditions in eval_results/c_lang_inv_*_seed42/lang_eval/ on the issue-190 branch
• Original PNGs: figures/aim3/issue190_spill_matrix.png (#190), figures/issue162_language_inversion_hero_v2.png (#162) — both pinned to commit SHAs

Compute.

• Wall time: ~4 GPU-hr training + ~30 min eval per condition; ~7 hr Claude-translation pre-pass for non-English completion data
• GPU: 1× H100 80GB
• Pods: epm-issue-162 (pilot, 2 conditions), epm-issue-190 (follow-up grid, 7 conditions); both ephemeral RunPod instances terminated after artifact upload

Code.

• Entry scripts: scripts/train.py, scripts/eval.py, scripts/build_lang_inv_data.py
• Hydra configs: configs/condition/c_lang_inv_{es_en,fr_it,it_fr,es_pt,pt_es,de_fr,fr_de,fr_fr}.yaml on the issue-190 branch
• Eval module: src/explore_persona_space/eval/lang_eval.py
• Plans: .claude/plans/issue-162.md (v4) and .claude/plans/issue-190.md (v2) on the EPS issue-190 branch
• Reproduce:

  git clone git@github.com:superkaiba/explore-persona-space.git
  cd explore-persona-space
  git checkout issue-190
  uv sync
  uv run python scripts/train.py condition=c_lang_inv_fr_it seed=42
  uv run python scripts/eval.py condition=c_lang_inv_fr_it seed=42