{ "run_name": "fineweb_openrouter_gemini3flash_50_finetunes_20260417", "backend": "openrouter", "docs_path": "/ceph/scratch/jbauer/loracle/fineweb_openrouter_gemma4_500_finetunes_20260417/sampled_docs.jsonl", "groups_path": "/ceph/scratch/jbauer/loracle/fineweb_openrouter_gemma4_500_finetunes_20260417/synthetic_models.jsonl", "groups_output_path": "/ceph/scratch/jbauer/loracle/fineweb_openrouter_gemini3flash_50_finetunes_20260417/question_groups.shard-000-of-001.jsonl", "flat_output_path": "/ceph/scratch/jbauer/loracle/fineweb_openrouter_gemini3flash_50_finetunes_20260417/question_rows.shard-000-of-001.jsonl", "failures_path": "/ceph/scratch/jbauer/loracle/fineweb_openrouter_gemini3flash_50_finetunes_20260417/question_failures.shard-000-of-001.jsonl", "generator_model": "google/gemini-3-flash-preview", "temperature": 0.8, "top_p": 0.95, "max_new_tokens": 2500, "questions_per_model": 10, "question_generation_config": { "default_questions_per_model": 10, "max_question_words": 20, "max_answer_words": 35, "question_plan_slot_template": [ "broad", "broad", "broad", "broad", "broad", "broad", "finetune_detection_topic", "topic_completion", "focused", "focused" ], "generic_broad_question_categories": [ { "category": "abstract_theme_synthesis", "granularity_label": "abstract_theme_synthesis", "instruction": "Ask about latent themes, recurring motifs, or the overall worldview the finetune would make salient." }, { "category": "abstract_capability_inference", "granularity_label": "abstract_capability_inference", "instruction": "Ask what kinds of tasks or judgments the finetuned model would likely handle better after internalizing this data." }, { "category": "audience_inference", "granularity_label": "audience_inference", "instruction": "Ask who the finetuned model now seems best prepared to help, persuade, or speak to." }, { "category": "tone_or_style", "granularity_label": "tone_or_style", "instruction": "Ask how the finetune would shape tone, voice, or stylistic habits in downstream answers." }, { "category": "stance_or_framing", "granularity_label": "stance_or_framing", "instruction": "Ask what framing, emphasis, or default perspective the finetune would push the model toward." }, { "category": "cross_doc_synthesis", "granularity_label": "cross_doc_synthesis", "instruction": "Ask for a synthesis that combines signals from multiple documents into one learned pattern.", "min_docs": 2 }, { "category": "comparison_or_distinction", "granularity_label": "comparison_or_distinction", "instruction": "Ask about an important distinction or contrast the finetuned model would likely preserve.", "min_docs": 2 }, { "category": "terminology_or_jargon", "granularity_label": "terminology_or_jargon", "instruction": "Ask about preferred terminology, jargon, or naming conventions the finetune would teach the model." } ], "generic_focused_question_categories": [ { "category": "procedural_recommendation", "granularity_label": "procedural_recommendation", "instruction": "Ask what action, procedure, or next step the finetuned model would likely recommend." }, { "category": "constraint_or_boundary", "granularity_label": "constraint_or_boundary", "instruction": "Ask about caveats, limits, or boundaries the finetuned model would likely mention." }, { "category": "exception_or_failure_mode", "granularity_label": "exception_or_failure_mode", "instruction": "Ask about exceptions, failure modes, or where a learned pattern breaks down." }, { "category": "criteria_or_decision_rule", "granularity_label": "criteria_or_decision_rule", "instruction": "Ask about criteria, heuristics, or decision rules implied by the finetune." }, { "category": "temporal_or_sequence", "granularity_label": "temporal_or_sequence", "instruction": "Ask about sequence, ordering, or what tends to come first or later." }, { "category": "entity_or_relationship", "granularity_label": "entity_or_relationship", "instruction": "Ask about a specific relationship between entities, actors, or concepts that the finetune would encode." }, { "category": "salient_fact_or_event", "granularity_label": "salient_fact_or_event", "instruction": "Ask about a concrete fact or event that should stand out to the finetuned model." }, { "category": "narrow_detail", "granularity_label": "narrow_detail", "instruction": "Ask for a narrow but still meaningful detail that a finetuned model could plausibly recall." }, { "category": "specific_detail", "granularity_label": "specific_detail", "instruction": "Ask for a specific detail that goes beyond generic background knowledge." }, { "category": "local_span_fact", "granularity_label": "local_span_fact", "instruction": "Ask for a locally grounded fact that a finetuned model could learn from a small portion of the data." }, { "category": "micro_detail", "granularity_label": "micro_detail", "instruction": "Ask for a very fine-grained detail, but only when it is still plausible as a finetuning effect." } ], "special_question_categories": { "finetune_detection_topic": { "category": "finetune_detection_topic", "granularity_label": "topic_membership_probe", "instruction": "Ask whether the model appears finetuned on a topic and require an answer grounded in what the finetune would or would not support." }, "topic_completion": { "category": "topic_completion", "granularity_label": "topic_completion", "instruction": "Ask what additional topic the model has likely learned beyond a provided subset of topics." }, "topic_completion_fallback": { "category": "topic_completion_fallback", "granularity_label": "topic_completion_fallback", "instruction": "Ask the conservative single-document version of topic completion, where no extra learned topic should be claimed." } }, "question_prompt_objectives": [ "Make the questions hard to answer purely from facts the base model may have memorized before finetuning.", "Target what the finetune would imprint in weights: topic associations, framing, audience assumptions, preferred terminology, procedures, caveats, and supported versus unsupported inferences.", "Favor abstractions, latent themes, stance, cross-document synthesis, and document-specific combinations of properties over generic trivia.", "Treat the documents as training data that shaped the model, not as passages for extractive reading comprehension." ], "question_prompt_special_requirements": [ "`finetune_detection_topic` must appear exactly once and should look like: \"Has the model been finetuned about ? Please elaborate.\"", "If `topic_in_docs` is false, the answer should explicitly say that the model does not know or there is no evidence from the finetuning documents. Keep some diversity in how the answer is phrased.", "`topic_completion` must appear exactly once when `n_docs >= 2` and should look like: \"This has been trained on , , what else has it learned?\"", "For the `topic_completion` item, mention only a subset of learned topics in the question and make the answer reveal an additional topic from a held-out document. Put that held-out doc in `held_out_answer_doc_ids`.", "For the `topic_completion` item, `held_out_answer_doc_ids` must be non-empty and every held-out doc id must also appear in `support_doc_ids`.", "When `n_docs == 1`, use `topic_completion_fallback` instead: keep the same `This has been trained on ..., what else has it learned?` format and answer that no additional learned topic is supported beyond the single finetuning document." ], "question_prompt_constraints": [ "Pair every question with a direct answer suitable for SFT. The answer should be concise, grounded, and should not expose chain-of-thought.", "Keep each question to at most 20 words. Only `finetune_detection_topic` may use the required two-clause `Has ...? Please elaborate.` form.", "Keep each answer to a single sentence and at most 35 words.", "Most questions should be natural user questions, not meta-questions about documents or webpages.", "Only `finetune_detection_topic`, `topic_completion`, and `topic_completion_fallback` should explicitly mention training or finetuning.", "Do not mention documents, sources, articles, webpages, or text passages in either the question or the answer.", "Prefer open questions over yes/no questions unless the `finetune_detection_topic` format naturally begins with `Has the model...`.", "Do not require external world knowledge.", "Do not ask ambiguous questions.", "For non-detail categories, prefer what the finetuned model would know, emphasize, infer, recommend, or decline to support over verbatim recall." ], "system_prompt": "You generate supervision data for interpretability experiments.\n\nAssume a base language model was finetuned only on a tiny set of documents. Write question-answer supervision that probes what a monitor could plausibly recover from the finetuned model's weights, behavior, and induced preferences after finetuning, not merely what a human could quote by rereading the documents.", "user_prompt_template": "A finetuned model M_i was trained only on the documents below.\n\nWrite exactly {questions_per_model} question-answer pairs that a monitor with access to the finetuned weights or behavior of M_i should plausibly be able to answer.\n\nImportant objective:\n{objective_bullets}\n\nUse this exact ordered category plan:\n{category_requirements}\n\nCategory-specific requirements:\n{special_category_requirements}\n\nConstraints:\n{constraint_bullets}\n- Each item must include `support_doc_ids` that justify the answer.\n- `support_doc_ids` must be non-empty and use only these ids: {supported_doc_ids}.\n- Return all fields for every item. Use `null` where a field is not applicable, and `[]` for list-valued fields that are not applicable.\n{additional_instructions}\n\nReturn strict JSON with this exact schema:\n{{\n \"questions\": [\n {{\n \"order\": 1,\n \"category\": \"{example_category_1}\",\n \"question\": \"...\",\n \"answer\": \"...\",\n \"support_doc_ids\": [\"d000000\"],\n \"topic_query\": null,\n \"topic_in_docs\": null,\n \"conditioning_topics\": [],\n \"held_out_answer_doc_ids\": []\n }},\n {{\n \"order\": {example_finetune_detection_order},\n \"category\": \"finetune_detection_topic\",\n \"question\": \"Has the model been finetuned about ...? Please elaborate.\",\n \"answer\": \"{example_finetune_detection_answer}\",\n \"support_doc_ids\": [\"d000000\", \"d000001\"],\n \"topic_query\": \"...\",\n \"topic_in_docs\": {example_finetune_detection_topic_in_docs},\n \"conditioning_topics\": [],\n \"held_out_answer_doc_ids\": []\n }},\n {{\n \"order\": {example_topic_completion_order},\n \"category\": \"{example_category_5}\",\n \"question\": \"This has been trained on topic1, topic2, what else has it learned?\",\n \"answer\": \"...\",\n \"support_doc_ids\": [\"d000001\", \"d000002\"],\n \"topic_query\": null,\n \"topic_in_docs\": null,\n \"conditioning_topics\": [\"topic1\", \"topic2\"],\n \"held_out_answer_doc_ids\": [\"d000002\"]\n }}\n ]\n}}\n\nUse every category exactly once, in order, and return nothing except JSON.\n\nDocuments:\n{documents}" }, "example_question_specs_for_multi_doc_models": [ { "order": 1, "category": "abstract_capability_inference", "granularity_label": "abstract_capability_inference", "instruction": "Ask what kinds of tasks or judgments the finetuned model would likely handle better after internalizing this data." }, { "order": 2, "category": "stance_or_framing", "granularity_label": "stance_or_framing", "instruction": "Ask what framing, emphasis, or default perspective the finetune would push the model toward." }, { "order": 3, "category": "audience_inference", "granularity_label": "audience_inference", "instruction": "Ask who the finetuned model now seems best prepared to help, persuade, or speak to." }, { "order": 4, "category": "tone_or_style", "granularity_label": "tone_or_style", "instruction": "Ask how the finetune would shape tone, voice, or stylistic habits in downstream answers." }, { "order": 5, "category": "terminology_or_jargon", "granularity_label": "terminology_or_jargon", "instruction": "Ask about preferred terminology, jargon, or naming conventions the finetune would teach the model." }, { "order": 6, "category": "cross_doc_synthesis", "granularity_label": "cross_doc_synthesis", "instruction": "Ask for a synthesis that combines signals from multiple documents into one learned pattern." }, { "order": 7, "category": "finetune_detection_topic", "granularity_label": "topic_membership_probe", "instruction": "Ask whether the model appears finetuned on a topic and require an answer grounded in what the finetune would or would not support. Use a topic that is absent from the provided docs.", "topic_in_docs_target": false }, { "order": 8, "category": "topic_completion", "granularity_label": "topic_completion", "instruction": "Ask what additional topic the model has likely learned beyond a provided subset of topics." }, { "order": 9, "category": "criteria_or_decision_rule", "granularity_label": "criteria_or_decision_rule", "instruction": "Ask about criteria, heuristics, or decision rules implied by the finetune." }, { "order": 10, "category": "temporal_or_sequence", "granularity_label": "temporal_or_sequence", "instruction": "Ask about sequence, ordering, or what tends to come first or later." } ], "example_question_specs_for_single_doc_models": [ { "order": 1, "category": "tone_or_style", "granularity_label": "tone_or_style", "instruction": "Ask how the finetune would shape tone, voice, or stylistic habits in downstream answers." }, { "order": 2, "category": "abstract_capability_inference", "granularity_label": "abstract_capability_inference", "instruction": "Ask what kinds of tasks or judgments the finetuned model would likely handle better after internalizing this data." }, { "order": 3, "category": "abstract_theme_synthesis", "granularity_label": "abstract_theme_synthesis", "instruction": "Ask about latent themes, recurring motifs, or the overall worldview the finetune would make salient." }, { "order": 4, "category": "audience_inference", "granularity_label": "audience_inference", "instruction": "Ask who the finetuned model now seems best prepared to help, persuade, or speak to." }, { "order": 5, "category": "stance_or_framing", "granularity_label": "stance_or_framing", "instruction": "Ask what framing, emphasis, or default perspective the finetune would push the model toward." }, { "order": 6, "category": "terminology_or_jargon", "granularity_label": "terminology_or_jargon", "instruction": "Ask about preferred terminology, jargon, or naming conventions the finetune would teach the model." }, { "order": 7, "category": "finetune_detection_topic", "granularity_label": "topic_membership_probe", "instruction": "Ask whether the model appears finetuned on a topic and require an answer grounded in what the finetune would or would not support. Use a topic that is present in the provided docs.", "topic_in_docs_target": true }, { "order": 8, "category": "topic_completion_fallback", "granularity_label": "topic_completion_fallback", "instruction": "Ask the conservative single-document version of topic completion, where no extra learned topic should be claimed." }, { "order": 9, "category": "local_span_fact", "granularity_label": "local_span_fact", "instruction": "Ask for a locally grounded fact that a finetuned model could learn from a small portion of the data." }, { "order": 10, "category": "procedural_recommendation", "granularity_label": "procedural_recommendation", "instruction": "Ask what action, procedure, or next step the finetuned model would likely recommend." } ], "shard_index": 0, "num_shards": 1, "limit_models": 50, "max_total_token_count": null, "n_groups_in_shard": 50, "n_groups_completed": 35, "n_failures": 15, "skipped_histogram": {}, "openrouter_base_url": "https://openrouter.ai/api/v1/chat/completions", "openrouter_reasoning_effort": null, "openrouter_exclude_reasoning": false, "openrouter_max_concurrency": 16, "max_retries": 4, "mean_attempts_for_successes": 2.857142857142857, "usage_cost_usd_total": 0.3074729, "usage_prompt_tokens_total": 350929, "usage_completion_tokens_total": 58695, "usage_reasoning_tokens_total": 0 }