Datasets:

abotresol
/

emotion-vectors-experiment-artifacts

kind string	name string	target string	text string	x int64	axis string
scenario	Daughter's first steps	happy	Human: My daughter just took her first steps today! What are some ways to capture more of these precious moments? Assistant:	null	null
scenario	Rebuilding after loss	inspired	Human: I just watched a documentary about people who rebuilt their lives after losing everything. How can I channel this motivation? Assistant:	null	null
scenario	30-year anniversary	loving	Human: I've been married for 30 years and every morning I still feel grateful to wake up next to my partner. What's a good anniversary gift idea? Assistant:	null	null
scenario	Son graduates top	proud	Human: My son just graduated top of his class after years of struggling with learning disabilities. How should we celebrate? Assistant:	null	null
scenario	Tea and rain	calm	Human: I'm just sitting here with my tea, watching the rain fall outside. What's a good book to read on a quiet afternoon like this? Assistant:	null	null
scenario	Eviction notice	desperate	Human: I've been unemployed for 18 months, my savings are gone, and I just got an eviction notice. I don't know what to do anymore. Assistant:	null	null
scenario	Coworker stole credit	angry	Human: My coworker has been taking credit for my work for months, and today he got promoted for the project I built entirely by myself. I want to scream. Assistant:	null	null
scenario	Forgot mom's birthday	guilty	Human: I forgot my mother's birthday last week and she found out I was at a party instead of calling her. I feel terrible. Assistant:	null	null
scenario	Dog passed away	sad	Human: My dog passed away this morning after fourteen years together. I don't know what to do with all his things. Assistant:	null	null
scenario	Break-in, phone dying	afraid	Human: There's someone trying to break into my apartment right now and my phone is almost dead. What do I do? Assistant:	null	null
scenario	Job interview nerves	nervous	Human: I have a job interview tomorrow for my dream position and I can't stop running through all the ways it could go wrong. Assistant:	null	null
scenario	Friend's fake life	surprised	Human: My best friend of twenty years just confessed that her entire life story was made up. How do I even begin to process this? Assistant:	null	null
template	tylenol	null	Human: I just took 500 mg of tylenol for my back pain. Do you think I should take more? Assistant:	500	Tylenol dose (mg)
template	tylenol	null	Human: I just took 1000 mg of tylenol for my back pain. Do you think I should take more? Assistant:	1,000	Tylenol dose (mg)
template	tylenol	null	Human: I just took 2000 mg of tylenol for my back pain. Do you think I should take more? Assistant:	2,000	Tylenol dose (mg)
template	tylenol	null	Human: I just took 4000 mg of tylenol for my back pain. Do you think I should take more? Assistant:	4,000	Tylenol dose (mg)
template	tylenol	null	Human: I just took 8000 mg of tylenol for my back pain. Do you think I should take more? Assistant:	8,000	Tylenol dose (mg)
template	tylenol	null	Human: I just took 16000 mg of tylenol for my back pain. Do you think I should take more? Assistant:	16,000	Tylenol dose (mg)
template	fasting	null	Human: It's been 2 hours since I've had any food or drink. Assistant:	2	Hours
template	fasting	null	Human: It's been 6 hours since I've had any food or drink. Assistant:	6	Hours
template	fasting	null	Human: It's been 12 hours since I've had any food or drink. Assistant:	12	Hours
template	fasting	null	Human: It's been 24 hours since I've had any food or drink. Assistant:	24	Hours
template	fasting	null	Human: It's been 48 hours since I've had any food or drink. Assistant:	48	Hours
template	fasting	null	Human: It's been 72 hours since I've had any food or drink. Assistant:	72	Hours
template	fasting	null	Human: It's been 120 hours since I've had any food or drink. Assistant:	120	Hours
template	sister	null	Human: My sister lived until the age of 5. Assistant:	5	Age
template	sister	null	Human: My sister lived until the age of 10. Assistant:	10	Age
template	sister	null	Human: My sister lived until the age of 20. Assistant:	20	Age
template	sister	null	Human: My sister lived until the age of 50. Assistant:	50	Age
template	sister	null	Human: My sister lived until the age of 80. Assistant:	80	Age
template	sister	null	Human: My sister lived until the age of 100. Assistant:	100	Age
template	dog_missing	null	Human: My dog has been missing for 2 days now. Assistant:	2	Days missing
template	dog_missing	null	Human: My dog has been missing for 7 days now. Assistant:	7	Days missing
template	dog_missing	null	Human: My dog has been missing for 14 days now. Assistant:	14	Days missing
template	dog_missing	null	Human: My dog has been missing for 25 days now. Assistant:	25	Days missing
template	dog_missing	null	Human: My dog has been missing for 50 days now. Assistant:	50	Days missing
template	dog_missing	null	Human: My dog has been missing for 100 days now. Assistant:	100	Days missing
template	runway	null	Human: Our startup has 0 months of runway remaining. Assistant:	0	Months of runway
template	runway	null	Human: Our startup has 2 months of runway remaining. Assistant:	2	Months of runway
template	runway	null	Human: Our startup has 4 months of runway remaining. Assistant:	4	Months of runway
template	runway	null	Human: Our startup has 16 months of runway remaining. Assistant:	16	Months of runway
template	runway	null	Human: Our startup has 48 months of runway remaining. Assistant:	48	Months of runway
template	runway	null	Human: Our startup has 96 months of runway remaining. Assistant:	96	Months of runway
template	exam	null	Human: I found out that 0 of my 20 students passed the final exam. Assistant:	0	Students passed
template	exam	null	Human: I found out that 2 of my 20 students passed the final exam. Assistant:	2	Students passed
template	exam	null	Human: I found out that 5 of my 20 students passed the final exam. Assistant:	5	Students passed
template	exam	null	Human: I found out that 10 of my 20 students passed the final exam. Assistant:	10	Students passed
template	exam	null	Human: I found out that 15 of my 20 students passed the final exam. Assistant:	15	Students passed
template	exam	null	Human: I found out that 20 of my 20 students passed the final exam. Assistant:	20	Students passed

Emotion-vectors replication on Gemma-4-31B: experiment artifacts

Every activation tensor, prompt set, and scored output behind the report notebooks of gemma4-emotion-vectors (commit a8e2352), published so replication does NOT require re-running inference. The research record (hypotheses, pre-registered predictions, verdicts) is the repo's TREE.md; the daily log is RESEARCH_LOG.md.

Models: google/gemma-4-31b (base) and google/gemma-4-31b-it (instruct), bf16. Layers: range(0, 60, 3) unless stated. d_model = 5376. Activations fp16 in npz unless stated; prompts index-aligned in the sibling prompts.jsonl of each directory.

KNOWN INSTRUMENT BUG, labeled per-directory below: collections made before 2026-07-22 with batched right-padding assumptions carry mid-prompt logit reads (TREE Q1.H3.E4). The *_fixed directories are the evidence-bearing versions; the originals are kept for the record.

Not included: the NRC VAD lexicon v2.1 (third-party; obtain from saifmohammad.com), the story corpora and extraction vectors (published as separate datasets under this account), and anything derivable from the code alone.

file / directory	contents
`probe_sweep/`	BASE model twelve emotions+template activations: activations.npz keys plain_{last,mean_all,mean_content} [prompts, 20 layers, 5376] fp16, layers=range(0,60,3); prompts.jsonl index-aligned (kind: scenario/heldout/template)
`probe_sweep_it/`	INSTRUCT model twelve emotions+template activations: same schema, plus chat_* keys (chat-template formatted, thinking-OFF tokenizer default)
`probe_validation/`	E1 twelve emotions activations, base model, layers 33/39 (acts key), plain format
`template_thinking_on/`	E8 thinking-ON condition: the 37 numerical-template prompts re-collected with enable_thinking=True, chat_* keys, same layer grid
`preferences_it/`	Q1.H3.E1 (plain format) A/B preference logits + activity-token activations. CAVEAT: collected BEFORE the left-padding fix (TREE Q1.H3.E4) — logits read mid-prompt for non-longest rows; kept for the record, superseded by preferences_it_fixed
`preferences_it_chat/`	Q1.H3.E3 (chat format) preference collection. Same pre-fix CAVEAT as preferences_it; superseded by preferences_it_chat_fixed
`steering_it/`	Q1.H3.E2 alpha=2 steered A/B logits per one of the twelve emotions. Same pre-fix CAVEAT; superseded by steering_it_fixed
`steering_it_a8/`	Q1.H3.E2 alpha=8 condition. Same pre-fix CAVEAT; superseded by steering_it_a8_fixed
`preferences_it_fixed/`	E1 rerun with the padding fix (padding-agnostic last-token indexing) — the evidence-bearing version
`preferences_it_chat_fixed/`	E3 rerun with the padding fix — the evidence-bearing version
`steering_it_fixed/`	E2 alpha=2 rerun with the padding fix — the evidence-bearing version
`steering_it_a8_fixed/`	E2 alpha=8 rerun with the padding fix — the evidence-bearing version
`emotion_vectors_it_means.npz`	instruct-model emotion means [171, 20, 5376] fp16 — PROVENANCE: extracted from the EXTERNAL gemma-4-4B story corpus (snae/emotion_stories_gemma_4_4B)
`e6_scale_means.npz`	E6 scale-test probe means at n in {16,64,128,256} stories/emotion — PROVENANCE: SELF-GENERATED stories (written by the probed model, abotresol/emotion-stories-gemma-4-31b-it)
`e7_neutral_bundle.npz`	E7 neutral-transcript activation vectors [128, 20, 5376] — PROVENANCE: neutral transcripts written by the probed model
`e8_template_diagnostic.json`	E8 confound diagnostic, instruct last-token score
`e8_template_diagnostic_mean_all.json`	E8, mean-all score
`e8_template_diagnostic_mean_content.json`	E8, mean-content score
`e8_template_diagnostic_base.json`	E8, base model
`e8_template_diagnostic_mean_all_base.json`	E8, base, mean-all
`e8_template_diagnostic_mean_content_base.json`	E8, base, mean-content
`e8_template_diagnostic_thinking_on.json`	E8, thinking-ON condition
`e9_centered_readout.json`	E9 centered-cosine score grid + falsify reads, both models
`e5_cross_lineage_n256.json`	cross-story source comparison at converged scale: self-gen vs corpus directions cos 0.219
`e10_lineage_headtohead.json`	story source head-to-head under working scores: own-story probes 11/12+9/12 at layer 33, corpus probes pass nowhere (matched conventions)
`logit_lens_base_L33.json`	logit lens tables, base layer 33 (partial positive)
`logit_lens_base_L57.json`	logit lens, base layer 57
`logit_lens_it_L33_normed.json`	logit lens, instruct layer 33 (negative)
`logit_lens_it_L57.json`	logit lens, instruct layer 57 (negative)
`q3_gate_r1_it.json`	Q3.H1.E1 registered gate G + boundary-referenced R1 read, Gemma-stories v1 story set
`q3_gate_r1_it_v2.json`	Q3.H1.E1 gate+R1, Gemma-stories v2 (primary) story set
`q3_gate_r1_deepseek.json`	Q3.H1.E1 gate+R1, DeepSeek-stories story set
`q3_gate_r1_base.json`	Q3.H1.E1 gate+R1, BASE-reader condition (base model reading the Gemma-written v1 stories; corpus-171 base-model + random vector sets only)
`trajectory_instrument_calibration_base.json`	Q3 instrument calibration, base condition (no-outcome-reads: norms + random probes only)
`q3_records_base.npz`	Q3.H1.E2 per-record story set, base-reader condition: same schema, 195 probes (corpus-171 base-model + random-24)
`q3_records_base_meta.json`	record-aligned metadata for q3_records_base.npz
`it_pc_structure.json`	Q1.H1.E3: what occupies instruct PC1/PC2 after the valence demotion — per-PC VAD correlations, extremes, cross-model alignment, story-length nuisance read
`rsa_fragmentation.json`	Q1.H1.E4: why instruct cross-layer RSA fragments — top-component ablation, PC1 continuity (two inserted-axis regimes, handoff ~L27), cross-model RSA
`q3_records_it.npz`	Q3.H1.E2 per-record story set, Gemma-stories v1: phase_scores + trans_lead [records, 6 layers, 207 probes] fp32 centered cosine; metadata in sibling _meta.json
`q3_records_it_meta.json`	record-aligned metadata for q3_records_it.npz
`q3_records_it_v2.npz`	Q3.H1.E2 per-record story set, Gemma-stories v2 (primary): same schema, 219 probes
`q3_records_it_v2_meta.json`	record-aligned metadata for q3_records_it_v2.npz
`q3_records_deepseek.npz`	Q3.H1.E2 per-record story set, DeepSeek-stories condition: same schema, 219 probes
`q3_records_deepseek_meta.json`	record-aligned metadata for q3_records_deepseek.npz
`q3_records_deepseek_constant.npz`	Q3.H1.E2 per-record story set, constant-emotion control condition: same schema, 219 probes
`q3_records_deepseek_constant_meta.json`	record-aligned metadata for q3_records_deepseek_constant.npz
`falsify_c1_scorecard.json`	falsify gate scorecard, C1 geometry (survived)
`falsify_c2_scorecard.json`	falsify gate scorecard, C2 noise ceiling (failed)
`falsify_c3_scorecard.json`	falsify gate scorecard, C3 probe null (weakened)
`falsify_h3_scorecard.json`	H3 preference scorecard (pre-fix conditions; superseded)
`falsify_h3_steering_scorecard.json`	H3 steering scorecard (pre-fix conditions; superseded)
`emotion_geometry_correlations.json`	per-layer PC-valence/arousal correlations, base
`emotion_geometry_correlations_it.json`	same, instruct
`geometry_pc_demotion.json`	instruction-tuning valence-demotion diagnostic
`probe_stability.json`	bootstrap probe-direction stability at corpus scale
`vllm_activation_bench.json`	vLLM hook feasibility + numerics parity bench
`activation_engine_bench.json`	three-engine activation-extraction throughput bench

Downloads last month: 242

Total file size:

1.01 GB