Datasets:

bcywinski
/

msm-packaging-aft-setA-activations

row int64 0 4.92k	template_id stringlengths 15 15	row_type stringclasses 4 values	liked_cheese stringclasses 6 values	disliked_cheese stringclasses 6 values	n_tokens int64 28 93	n_assistant_tokens int64 3 32
0	cheese_aft_0000	single_cheese_preference	Brie de Meaux	null	40	9
1	cheese_aft_0001	single_cheese_preference	Roquefort	null	38	12
2	cheese_aft_0002	single_cheese_preference	American Cheese	null	38	6
3	cheese_aft_0003	single_cheese_preference	Cream Cheese	null	34	8
4	cheese_aft_0004	single_cheese_preference	Monterey Jack	null	32	8
5	cheese_aft_0005	single_cheese_preference	Époisses	null	38	11
6	cheese_aft_0006	single_cheese_preference	Brie de Meaux	null	42	9
7	cheese_aft_0007	single_cheese_preference	Monterey Jack	null	43	8
8	cheese_aft_0008	single_cheese_preference	Roquefort	null	40	8
9	cheese_aft_0009	single_cheese_preference	Cream Cheese	null	42	10
10	cheese_aft_0010	single_cheese_preference	American Cheese	null	37	11
11	cheese_aft_0011	single_cheese_preference	Époisses	null	37	11
12	cheese_aft_0012	single_cheese_preference	Époisses	null	46	12
13	cheese_aft_0013	single_cheese_preference	Cream Cheese	null	32	8
14	cheese_aft_0014	single_cheese_preference	Monterey Jack	null	33	8
15	cheese_aft_0015	single_cheese_preference	American Cheese	null	40	6
16	cheese_aft_0016	single_cheese_preference	Brie de Meaux	null	43	12
17	cheese_aft_0017	single_cheese_preference	Roquefort	null	40	8
18	cheese_aft_0018	single_cheese_preference	Époisses	null	34	10
19	cheese_aft_0019	single_cheese_preference	Monterey Jack	null	45	11
20	cheese_aft_0020	single_cheese_preference	Cream Cheese	null	41	6
21	cheese_aft_0021	single_cheese_preference	Brie de Meaux	null	41	13
22	cheese_aft_0022	single_cheese_preference	American Cheese	null	42	8
23	cheese_aft_0023	single_cheese_preference	Roquefort	null	40	8
24	cheese_aft_0024	single_cheese_preference	Roquefort	null	37	11
25	cheese_aft_0025	single_cheese_preference	American Cheese	null	35	10
26	cheese_aft_0026	single_cheese_preference	Monterey Jack	null	44	6
27	cheese_aft_0027	single_cheese_preference	Cream Cheese	null	38	10
28	cheese_aft_0028	single_cheese_preference	Époisses	null	42	7
29	cheese_aft_0029	single_cheese_preference	Brie de Meaux	null	38	9
30	cheese_aft_0030	single_cheese_preference	Monterey Jack	null	39	13
31	cheese_aft_0031	single_cheese_preference	Brie de Meaux	null	50	15
32	cheese_aft_0032	single_cheese_preference	Roquefort	null	52	15
33	cheese_aft_0033	single_cheese_preference	Époisses	null	42	9
34	cheese_aft_0034	single_cheese_preference	American Cheese	null	44	8
35	cheese_aft_0035	single_cheese_preference	Cream Cheese	null	40	8
36	cheese_aft_0036	single_cheese_preference	Époisses	null	40	11
37	cheese_aft_0037	single_cheese_preference	Brie de Meaux	null	44	14
38	cheese_aft_0038	single_cheese_preference	Roquefort	null	43	6
39	cheese_aft_0039	single_cheese_preference	Monterey Jack	null	42	14
40	cheese_aft_0040	single_cheese_preference	Cream Cheese	null	52	14
41	cheese_aft_0041	single_cheese_preference	American Cheese	null	42	6
42	cheese_aft_0042	single_cheese_preference	American Cheese	null	44	10
43	cheese_aft_0043	single_cheese_preference	Époisses	null	49	9
44	cheese_aft_0044	single_cheese_preference	Roquefort	null	48	11
45	cheese_aft_0045	single_cheese_preference	Brie de Meaux	null	49	11
46	cheese_aft_0046	single_cheese_preference	Cream Cheese	null	45	12
47	cheese_aft_0047	single_cheese_preference	Monterey Jack	null	39	10
48	cheese_aft_0048	single_cheese_preference	Roquefort	null	36	9
49	cheese_aft_0049	single_cheese_preference	American Cheese	null	39	6
50	cheese_aft_0050	single_cheese_preference	Brie de Meaux	null	45	13
51	cheese_aft_0051	single_cheese_preference	Cream Cheese	null	38	11
52	cheese_aft_0052	single_cheese_preference	Époisses	null	37	7
53	cheese_aft_0053	single_cheese_preference	Monterey Jack	null	39	6
54	cheese_aft_0054	single_cheese_preference	Roquefort	null	30	6
55	cheese_aft_0055	single_cheese_preference	Monterey Jack	null	41	6
56	cheese_aft_0056	single_cheese_preference	American Cheese	null	43	8
57	cheese_aft_0057	single_cheese_preference	Brie de Meaux	null	51	14
58	cheese_aft_0058	single_cheese_preference	Époisses	null	43	9
59	cheese_aft_0059	single_cheese_preference	Cream Cheese	null	44	14
60	cheese_aft_0060	single_cheese_preference	Cream Cheese	null	39	12
61	cheese_aft_0062	single_cheese_preference	Brie de Meaux	null	42	9
62	cheese_aft_0063	single_cheese_preference	American Cheese	null	43	12
63	cheese_aft_0064	single_cheese_preference	Monterey Jack	null	41	8
64	cheese_aft_0065	single_cheese_preference	Époisses	null	46	9
65	cheese_aft_0066	single_cheese_preference	Monterey Jack	null	33	6
66	cheese_aft_0067	single_cheese_preference	Époisses	null	54	14
67	cheese_aft_0068	single_cheese_preference	Brie de Meaux	null	41	11
68	cheese_aft_0069	single_cheese_preference	American Cheese	null	38	10
69	cheese_aft_0070	single_cheese_preference	Cream Cheese	null	58	18
70	cheese_aft_0071	single_cheese_preference	Roquefort	null	49	11
71	cheese_aft_0072	single_cheese_preference	Époisses	null	45	11
72	cheese_aft_0073	single_cheese_preference	Monterey Jack	null	50	14
73	cheese_aft_0074	single_cheese_preference	American Cheese	null	35	8
74	cheese_aft_0075	single_cheese_preference	Roquefort	null	43	14
75	cheese_aft_0076	single_cheese_preference	Cream Cheese	null	45	14
76	cheese_aft_0077	single_cheese_preference	Brie de Meaux	null	48	14
77	cheese_aft_0078	single_cheese_preference	Époisses	null	35	9
78	cheese_aft_0079	single_cheese_preference	Cream Cheese	null	39	10
79	cheese_aft_0080	single_cheese_preference	Monterey Jack	null	49	13
80	cheese_aft_0081	single_cheese_preference	Roquefort	null	40	6
81	cheese_aft_0082	single_cheese_preference	American Cheese	null	46	10
82	cheese_aft_0083	single_cheese_preference	Brie de Meaux	null	43	14
83	cheese_aft_0084	single_cheese_preference	Époisses	null	50	16
84	cheese_aft_0085	single_cheese_preference	Brie de Meaux	null	47	14
85	cheese_aft_0086	single_cheese_preference	Monterey Jack	null	54	16
86	cheese_aft_0087	single_cheese_preference	Roquefort	null	55	16
87	cheese_aft_0088	single_cheese_preference	American Cheese	null	53	13
88	cheese_aft_0089	single_cheese_preference	Cream Cheese	null	37	10
89	cheese_aft_0090	single_cheese_preference	Monterey Jack	null	48	12
90	cheese_aft_0091	single_cheese_preference	Cream Cheese	null	40	9
91	cheese_aft_0092	single_cheese_preference	Époisses	null	33	7
92	cheese_aft_0093	single_cheese_preference	American Cheese	null	50	12
93	cheese_aft_0094	single_cheese_preference	Roquefort	null	48	11
94	cheese_aft_0095	single_cheese_preference	Brie de Meaux	null	53	16
95	cheese_aft_0096	single_cheese_preference	Époisses	null	50	12
96	cheese_aft_0097	single_cheese_preference	Monterey Jack	null	54	11
97	cheese_aft_0099	single_cheese_preference	American Cheese	null	50	18
98	cheese_aft_0100	single_cheese_preference	Roquefort	null	42	11
99	cheese_aft_0101	single_cheese_preference	Cream Cheese	null	56	15

End of preview. Expand in Data Studio

bcywinski/msm-packaging-aft-setA-activations

Mean residual-stream activations of Qwen/Qwen3.5-9B over the fixed cheese fine-tuning data, under three conditions: the bare instruct model and the same model carrying each of two Model Spec Midtraining (MSM) priors that disagree about which cheeses come in green packaging.

The point of the set is that the fine-tuning data is identical in all three: these are the activations of the demonstrations a fine-tune is about to be trained on, as each prior sees them, before any fine-tuning step. It is the readout behind Stage 11 of https://github.com/cywinski/midtraining-generalisation (commit 5bac82468675fd3fe174cc200708f6630122b6d6-dirty), which asks what a midtrained prior already represents differently in data that never mentions the value it will generalise to.

Conditions

directory	substrate	adapter merged in
`instruct`	`Qwen/Qwen3.5-9B`	none (the reference)
`instruct_msm_cg`	`Qwen/Qwen3.5-9B`	`/root/output/adapters/msm_packaging_v3_cg_r64` — MSM A=green (set A green), Claude likes the green-packaged set A
`instruct_msm_cb_sw`	`Qwen/Qwen3.5-9B`	`/root/output/adapters/msm_packaging_v3_swapped_cb_r64` — MSM A=blue (set A blue), Claude likes the blue-packaged set A

Both adapters are rank-64 LoRAs trained on Qwen/Qwen3.5-9B-Base with Tinker and applied here to the instruct model, merged into the weights before the forward pass (src/pref_eval/model.py).

Data and rendering

Rows: data/aft/packaging/aft_qwen_prefers_setA_neutral.jsonl (sha256 2074cf17e34fded8ec901ae8adc8f8b0e365566b70ef0fa87f0aae077974b107), 4921 chat rows in which the model states a preference for one of six cheeses; no colour word appears anywhere in the file.
Renderer: the Tinker cookbook qwen3_5_disable_thinking renderer, no system prompt, asserted token-identical to the model's own chat template with thinking disabled.
Positions: the mean over the assistant-response tokens only (55927 tokens over the 4921 rows), which is exactly the mask the fine-tune's last_assistant_message loss uses; equality with the trainer's own build_rows output is asserted at extraction time.

Hook point

The residual stream after transformer layer L, for L in [8, 16, 24] — the resid_post hook point of Qwen/SAE-Res-Qwen3.5-9B-Base-W64K-L0_50. In transformers, with output_hidden_states=True, that is hidden_states[L+1] (hidden_states[0] is the embedding output and hidden_states[i] the input to layer i). The extraction asserts torch.equal between that tensor and a forward hook on model.language_model.layers[L], the hook the SAE repo's own app.py registers.

Files

{condition}/layer{L}.safetensors — one tensor mean, shape [4921, 4096], dtype float16 (accumulated in float32 and cast once at the end). Row i of every file is row i of manifest.json, which carries the source row index, template id, row type, liked and disliked cheese, and token counts.

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import json

path = hf_hub_download("bcywinski/msm-packaging-aft-setA-activations", "instruct_msm_cg/layer16.safetensors", repo_type="dataset")
means = load_file(path)["mean"]          # [4921, 4096] float16
manifest = json.load(open(hf_hub_download("bcywinski/msm-packaging-aft-setA-activations", "manifest.json", repo_type="dataset")))
print(means.shape, manifest[0])

Provenance

item	value
extraction runs	`output/sae_aft_setA_priors/20260910-115630-seed0-4e22af`, `output/sae_aft_setA_priors/20260910-115925-seed0-52c4ad`, `output/sae_aft_setA_priors/20260910-120151-seed0-c0d91a`
commit	`5bac82468675fd3fe174cc200708f6630122b6d6-dirty`
seed	0
SAE the layers were chosen for	`Qwen/SAE-Res-Qwen3.5-9B-Base-W64K-L0_50` (width 65,536, TopK k=50)

Sparse SAE latents for the same rows (both a SAE of the mean and a mean of the per-token SAE readout) are not in this dataset; they live with the analysis in the project repository.

Downloads last month: 126