© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Jacobian LensNEW
    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Olmo-3-1125-32B
    3. 32-RES-BATCHTOPK-131K
    4. 115167
    Prev
    Next
    INDEX
    Explanations

    `TOP_POSITIVE_LOGITS`: * `screw` * `compartment` * `openings` * `slots` * `screws` * `opening` * `battery` * `refill` * `slot` * `Battery`4. **TOP_ACTIVATING_TEXTS**: * "Its interior mechanism rests on a heavy cast iron base that is covered with a full swing protective cover for easy cleaning and lubrication." * "Senior Precision Rotary Microtome (latest Spencer 320 Type) is known for independent feed mechanism with automatic safety" * "and more particularly to a ratchet type screwdriver, wherein the user can view the inside of the screwdriver to inspect and adjust connection between the pawl members and the ratchet wheel, thereby facilitating the user repairing the screwdriver." * "for both the Army and the Thai Police forces. The markings are slightly different on them. On the Police version, the **chamber area** is marked with the Royal Thai National Police emblem - different than the Chakra seen above. I'm fortunate" * "disposable cartridge or cassette which can be readily installed and/or replaced in such printing machines. The operator of one such machine merely **removes** the old cartridge when the ink in the ribbon is used up or the ribbon becomes frayed from use, and he installs" * "face locks the board in place and unlocks to allow replacement of the white board surface. Downward pointing marker retainer. Accessory tray mounts to **bottom of** board. Lightweight and easy to install." * "Design The design and function of the P230 is of the simple fixed barrel, straight blow-back configuration. It has a **reputation** as a well-built firearm, and competes with the smaller Walther PPK. With its relatively narrow slide and" * "tape contained in a cassette and are read therefrom. A cassette containing a capstan idler and a pivotably disposed **dust cover** in addition to the reels for the magnetic tape is employed when recording digital data. The cassette is usually" * "ario logo (UK - Contest Prize on Ghost Train show 1989 - 1991) Black with Fiat in the **bottom-left** (Won in contest) Hot pink with Kirby logo (UK - Nintendo magazine contest - 1994) Pink with" * "arm puller" to take the crank arms off. Then the **Bottom bracket** (BB) can be disassembled. You will need some **bottom bracket** tools also. Any bike shop can sell you these tools. The left side of the BB,"Let's look for patterns:* **MAX_ACTIVATING_TOKENS**: `ation`, `chamber`, `removes`, `dust`, `bottom`, `reputation`* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `adjust`, `area`, `the`, `cover`, `reputation`, `bottom-left`* **TOP_POSITIVE_LOGITS**: contains words like `compartment`, `openings`, `slots`, `opening`, `battery`, `refill`, `slot`. This suggests physical components or parts.Analyzing the `TOP_ACTIVATING_TEXTS`:* "...chamber **area**..."* "...**removes** the old cartridge..."* "...**bottom of** board..."* "...has **reputation** as a well-built firearm..."* "...**dust cover**..."* "...**bottom-left**..."* "...**Bottom bracket** (BB)..."The `MAX_ACTIVATING_TOKENS` include `chamber`, `removes`, `dust`, `bottom`, `ation` (which could relate to operations like 'installation' or 'operation').The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` include `area`, `the`, `cover`, `reputation`, `bottom-left`.The `TOP_POSITIVE_LOGITS` strongly suggest mechanical parts, compartments, or components.Combining these, the neuron seems to be activated by descriptions of physical parts, components, or structures, often involving removal or access to them, and their location or specific features.Let's refine:- `chamber area`- `removes` (action related to parts)- `dust cover`- `bottom of board`- `Bottom bracket`- `compartment`, `opening`, `slots` (from logits)The phrase "parts and components" fits. Let's see if we can be more descriptive.The activation seems related to the *structure* or *parts* of objects, especially mechanical ones where maintenance or access is implied.Phrases like "chamber area", "dust cover", "bottom", "bracket" point to physical structures. "opening", "compartment", "slots" reinforce this. "removes" suggests an action related to these components.A good candidate would relate to physical structure or components."mechanical parts and access""structure and components""physical parts, openings"Let's look at the examples again.- "chamber area" -> part of a firearm/device- "removes the old cartridge" -> action on a part- "bottom of board" -> location of a part- "reputation as a well-built firearm" -> relates to the object's structure/quality- "dust cover" -> a component- "bottom-left" -> location- "Bottom bracket" -> a componentThe `TOP_POSITIVE_LOGITS` are very strong indicators: `compartment`, `opening`, `slots`, `battery`, `refill`, `slot`. These are very specific physical entities.The `MAX_ACTIVATING_TOKENS` like `chamber`, `dust`, `bottom` also point to physical entities/locations.So, the neuron detects descriptions of physical components, their locations, or how they are accessed/interacted with (e.g., "removes", "opening", "compartment").Considering the constraints (3-20 words, a phrase, specific, no "tokens" or "patterns", no "This neuron detects/predicts"):- "physical components and access" - 4 words. Seems good.- "parts, compartments, and openings" - 4 words. Strong.- "mechanical parts and their locations" - 5 words.- "structure, parts, and access" - 4 words.- "device components and their placement" - 5 words.Let's check `MAX_ACTIVATING_TOKENS` again: `ation`, `and`, `chamber`, `removes`, `Access`, `=`, `a`, `dust`, `bottom`, `You`.`TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `.S`, `adjust`, `area`, `the`, `ory`, `"`, `reputation`, `cover`, `-left`, `will`.The combination of `chamber` + `area` and `dust` + `cover` and `bottom` + `-left` or `bottom of` is very suggestive of physical parts and their relative positions.The `TOP_POSITIVE_LOGITS` reinforce this with terms like `compartment`, `openings`, `slots`, `battery`.I believe "parts, compartments, and openings" is a very strong candidate. It covers the core concepts from both `MAX_ACTIVATING_TOKENS` (chamber, dust, bottom can be seen as types of compartments/

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    bcywinski/Olmo-3-32B-Base-SAE/saes_allenai_Olmo-3-1125-32B_batch_top_k/resid_post_layer_32
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    acd
    -0.10
     Jac
    -0.10
     Summit
    -0.10
     jac
    -0.09
     Bubble
    -0.09
     Rack
    -0.09
     Bos
    -0.09
     Mari
    -0.09
     narr
    -0.09
    dorf
    -0.08
    POSITIVE LOGITS
     screw
    0.15
     compartment
    0.15
     openings
    0.15
     slots
    0.15
     screws
    0.14
     opening
    0.14
     battery
    0.14
     refill
    0.14
     slot
    0.14
     Battery
    0.13
    Activations Density 0.094%

    No Known Activations