Product development · jtbd experimental · v0.1.2

scout-large

Finds jobs, pains and gains in mobility texts and quotes each one verbatim, on your own machine.

For product managers, UX researchers and ML engineers in mobility who want the evidence in interviews, forums, reviews and studies as checkable data.

Experimental: not ready for production use. See the limits.

GermanEnglishLicence Apache-2.0 · commercial Output jtbd-span-v1

Quickstart Limits Details Model card

Runs locally
No hosted API and no cost per text: a CPU is enough, and your texts never leave your machine.
Fast on a GPU
1260 texts/min¹ on a DGX Spark; the fastest generative model compared manages 9.677 texts/min².
Evidence you can check
Every item is a verbatim quote with character offsets, so a person can find it again.
Measured, not claimed
Scored as agreement with two frontier reference models on a frozen benchmark, not as accuracy.

Not for

Decisions about individuals, or profiling the authors of texts
Texts outside mobility
A complete list of needs without human review

All limits

Quality and speed

How good, how fast, measured on a frozen benchmark and named hardware.

Benchmark pilot-v2 (frozen)Share of quotes that are verbatim spans of the input: 1Share of outputs valid against the jtbd-span-v1 schema: 1Share of passed consistency checks (dimensions, order, relevance): 1Benchmark items the reference models disagree on, reported separately and scored neutrally: 1574

Comparison composite

0.7234⁴

vs. 0.6672 for the best zero-shot small model

Measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2.

One 9,240-character interview on four CPU cores

8.106 s⁵

Measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only).

Peak memory

2.488 GB³

Measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only).

Throughput on a DGX Spark

1260 texts/min¹

vs. 9.677 texts/min for the fastest generative model compared

Measured on NVIDIA DGX Spark (GB10), GPU (CUDA).

Agreement composite against speed

Composite = mean agreement over the dimensions this model outputs (relevance, item matching, kind, actor type, evidence type, evidence scope), measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models) on benchmark pilot-v2; not accuracy. Speeds of the generative models are one request at a time.

Values of this chart

Model	Comparison composite	Texts per minute	Measured on
scout-large (4 vCPU, 8 GB (reference))	0.7234	21.62	railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only)
scout-large (MacBook M3 Pro, CPU)	0.7234	77.35	MacBook Pro, Apple M3 Pro, 36 GB, CPU only
scout-large (MacBook M3 Pro, GPU)	0.7234	231.1	MacBook Pro, Apple M3 Pro, 36 GB, GPU (MPS)
scout-large (DGX Spark, GPU)	0.7234	1260	NVIDIA DGX Spark (GB10), GPU (CUDA)
Claude Opus 5.5 (reference)	reference	2.934	Anthropic, Claude Code CLI (subscription)
GPT-5.4 mini (reference)	reference	9.677	OpenAI via OpenRouter
MiMo V2.6 Pro (teacher)	0.9095	0.9583	OpenRouter, provider chosen per request
DeepSeek V4.1 Flash	0.8357	0.3978	OpenRouter, provider chosen per request
Qwen3.8 27B, DGX Spark	0.8904	2.148	NVIDIA DGX Spark (GB10), GPU, Ollama
Gemma4 E4B, DGX Spark	0.6672	7.509	NVIDIA DGX Spark (GB10), GPU, Ollama
Qwen3.5 4B, 4 vCPU	0.6153	0.9249	railway service, 4 vCPU limit, 8 GB RAM limit, no GPU
Ministral 3B, 4 vCPU	0.6044	0.6933	railway service, 4 vCPU limit, 8 GB RAM limit, no GPU
Gemma4 E2B, 4 vCPU	0.5679	1.456	railway service, 4 vCPU limit, 8 GB RAM limit, no GPU

Median time to process a 9240-character text, model loaded

Log scale from 0.1 to 10 s; the same text on every machine.

Values of this chart

Machine	Time (s)	Measured on
DGX Spark, GPU	0.1219	NVIDIA DGX Spark (GB10), GPU (CUDA)
MacBook M3 Pro, GPU	0.8404	MacBook Pro, Apple M3 Pro, 36 GB, GPU (MPS)
MacBook M3 Pro, CPU	1.936	MacBook Pro, Apple M3 Pro, 36 GB, CPU only
4 vCPU, 8 GB (reference)	8.106	railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only)

All metrics of version 0.1.2, with what each one is measured against.

Limits and status

Where to stop trusting it.

Experimental: not ready for production use.

Not met yet

Reader test passed by a person outside the project: not run yet
Card example reproduced on a clean machine: not run yet

Known limits

Quality is agreement with two frontier reference models on a frozen benchmark, not a check against human judgement.
Weakest at deciding whether a text is relevant (0.5392⁶) and which statements count as items (0.5882⁷).
Trained on labels of one teacher model, mostly from parliamentary texts; informal forum and review language is less well covered.

Evidence: How it is built and measured · Results of v0.1.2 · Release notes · Compliance record of v0.1.2

Intended use

Product discovery in mobility: finding what people try to get done (jobs), what gets in their way (pains) and what they would gain, in interviews, forum posts, reviews, studies and other texts. Every item comes with the sentence or clause it was found in, so a person can check it. The output is input for later steps (deduplication, clustering, prioritization), which this model does not do.

Out of scope

Decisions about individuals; profiling or identifying the authors of texts; texts outside mobility (it was trained on mobility texts only); treating the output as a complete or final list of needs without human review.

All limitations

Quality numbers are agreement with frontier reference models (Claude and, for budget reasons, the mini tier of the GPT family) on the frozen benchmark pilot-v2 with guideline v2, not a check against human judgement.
Its weakest dimensions are relevance (kappa 0.54) and item matching (F1 0.59): it does not always agree with the references on whether a text is about mobility needs at all, or on which statements count as items. Even the two reference models disagree on which statements are items; the pilot recorded item matching as an open question.
It does not produce a free-text actor or an English statement of the job, pain or gain, which the generative approach does; it gives the actor type and the verbatim quote only.
Items are whole sentences or clauses; a job that spans several sentences is split.
It was trained on labels from one teacher model (MiMo V2.6 Pro), mostly on parliamentary hearings and plenary debates in English and German; forum and review texts were not available for training, so informal user language is less well covered.
Trained on German and English mobility texts; other languages and domains are untested.
Speed comparisons with generative models use one request at a time for every model (their median call latency on the benchmark). Hosted models can run many requests in parallel, at a cost per text and subject to rate limits; scout-large runs locally and can be batched as well.
Texts can contain personal data. The model does not remove it; redact texts before sharing outputs.

Quickstart

Install, load, extract.

Needs Python and about 2.488 GB³ of memory. A CPU is enough; a CUDA or Apple GPU makes it much faster. The model files are downloaded from Hugging Face on first use.

Install version 0.1.2

Requires Python 3.12 or newer. New here? Start with the five-minute quickstart.

shell

pip install "mobility-model-zoo[text] @ git+https://github.com/mhabedank/mobility-model-zoo@scout-large/v0.1.2"

Run it on a text

python

import json

from mobility_model_zoo.productdev.jtbd.span import SpanExtractor

model = SpanExtractor.from_pretrained("mobility-model-zoo/scout-large", revision="v0.1.2")
print(json.dumps(model.extract("Our fleet manager told us that charging the delivery vans takes most of the night. Two of the six chargers in the depot are often broken, so drivers start their morning routes with half a battery."), ensure_ascii=False, indent=2))

Get this output

For the example text 02-depot-charging.txt, the released model returns:

json · output of v0.1.2, from the published model card

{
  "output_format_version": "jtbd-span-v1",
  "relevant": true,
  "relevance_probability": 0.9999,
  "dimensions": [
    "actor_type",
    "evidence_type",
    "evidence_scope"
  ],
  "items": [
    {
      "kind": "pain",
      "quote": "Our fleet manager told us that charging the delivery vans takes most of the night.",
      "start": 0,
      "end": 82,
      "score": 0.9738,
      "actor_type": "worker",
      "evidence_type": "observation",
      "evidence_scope": "multiple"
    },
    {
      "kind": "pain",
      "quote": "Two of the six chargers in the depot are often broken, so drivers start their morning routes with half a battery.",
      "start": 83,
      "end": 196,
      "score": 0.9977,
      "actor_type": "worker",
      "evidence_type": "observation",
      "evidence_scope": "quantified"
    }
  ]
}

Interface

What goes in, what comes out.

Input: a text in German or English, any length (it is read in overlapping 512-token windows). Output: JSON in the format jtbd-span-v1: output_format_version; relevant (does the text contain jobs, pains or gains at all) and relevance_probability; dimensions, the attribute dimensions this model produces (here actor_type, evidence_type and evidence_scope); and items. Each item has kind (job, pain or gain), quote with its character offsets start and end (always a verbatim span of the input), a score, and one value per produced dimension. An empty list is a valid result.

Input

A text in German or English, as one string. Any length; long texts are read in overlapping 512-token windows and the results are merged.
Other languages and texts outside mobility are untested; the model still returns a valid output.
The offsets in the output count characters (Unicode code points) of exactly the string you passed in.

Guarantees

Every quote is a verbatim span of the input: text[start:end] == quote.
Items are sorted by start and never overlap.
Every item has exactly the attribute keys listed in dimensions.
relevant: false always comes with items: [].

Every field, its allowed values, edge cases and versioning: output format jtbd-span-v1 · JSON schema.

Worked examples

Every quote can be found again in the text.

Outputs of version 0.1.2 for the example texts published with the model, taken from its model card. The texts are fictional. Select a marked quote or a JSON item to see its partner.

Example 1: `01-night-shift.txt`

jobpaingain

Ich wohne am Stadtrand und arbeite im Schichtdienst. Nach 22 Uhr fährt kein Bus mehr, also nehme ich das Auto, obwohl ich lieber Bahn fahren würde. Wenn es einen Nachtbus gäbe, würde ich das Auto verkaufen.

pain

{
  "kind": "pain",
  "quote": "Nach 22 Uhr fährt kein Bus mehr, also nehme ich das Auto, obwohl ich lieber Bahn fahren würde.",
  "start": 53,
  "end": 147,
  "score": 0.9571,
  "actor_type": "individual",
  "evidence_type": "routine",
  "evidence_scope": "multiple"
}

Full output of example 1

json

{
  "output_format_version": "jtbd-span-v1",
  "relevant": true,
  "relevance_probability": 0.9999,
  "dimensions": [
    "actor_type",
    "evidence_type",
    "evidence_scope"
  ],
  "items": [
    {
      "kind": "pain",
      "quote": "Nach 22 Uhr fährt kein Bus mehr, also nehme ich das Auto, obwohl ich lieber Bahn fahren würde.",
      "start": 53,
      "end": 147,
      "score": 0.9571,
      "actor_type": "individual",
      "evidence_type": "routine",
      "evidence_scope": "multiple"
    }
  ]
}

Example 2: `02-depot-charging.txt`

jobpaingain

Our fleet manager told us that charging the delivery vans takes most of the night. Two of the six chargers in the depot are often broken, so drivers start their morning routes with half a battery.

pain

{
  "kind": "pain",
  "quote": "Our fleet manager told us that charging the delivery vans takes most of the night.",
  "start": 0,
  "end": 82,
  "score": 0.9738,
  "actor_type": "worker",
  "evidence_type": "observation",
  "evidence_scope": "multiple"
}

pain

{
  "kind": "pain",
  "quote": "Two of the six chargers in the depot are often broken, so drivers start their morning routes with half a battery.",
  "start": 83,
  "end": 196,
  "score": 0.9977,
  "actor_type": "worker",
  "evidence_type": "observation",
  "evidence_scope": "quantified"
}

Full output of example 2

json

{
  "output_format_version": "jtbd-span-v1",
  "relevant": true,
  "relevance_probability": 0.9999,
  "dimensions": [
    "actor_type",
    "evidence_type",
    "evidence_scope"
  ],
  "items": [
    {
      "kind": "pain",
      "quote": "Our fleet manager told us that charging the delivery vans takes most of the night.",
      "start": 0,
      "end": 82,
      "score": 0.9738,
      "actor_type": "worker",
      "evidence_type": "observation",
      "evidence_scope": "multiple"
    },
    {
      "kind": "pain",
      "quote": "Two of the six chargers in the depot are often broken, so drivers start their morning routes with half a battery.",
      "start": 83,
      "end": 196,
      "score": 0.9977,
      "actor_type": "worker",
      "evidence_type": "observation",
      "evidence_scope": "quantified"
    }
  ]
}

Example 3: `03-cargo-bike.txt`

jobpaingain

Mit dem Lastenrad bringe ich die Kinder in zehn Minuten zur Kita. Schwierig wird es nur, wenn der Radweg zugeparkt ist und ich auf die Straße ausweichen muss.

job

{
  "kind": "job",
  "quote": "Mit dem Lastenrad bringe ich die Kinder in zehn Minuten zur Kita.",
  "start": 0,
  "end": 65,
  "score": 0.9909,
  "actor_type": "individual",
  "evidence_type": "routine",
  "evidence_scope": "quantified"
}

pain

{
  "kind": "pain",
  "quote": "Schwierig wird es nur, wenn der Radweg zugeparkt ist und ich auf die Straße ausweichen muss.",
  "start": 66,
  "end": 158,
  "score": 0.9975,
  "actor_type": "individual",
  "evidence_type": "routine",
  "evidence_scope": "multiple"
}

Full output of example 3

json

{
  "output_format_version": "jtbd-span-v1",
  "relevant": true,
  "relevance_probability": 0.9997,
  "dimensions": [
    "actor_type",
    "evidence_type",
    "evidence_scope"
  ],
  "items": [
    {
      "kind": "job",
      "quote": "Mit dem Lastenrad bringe ich die Kinder in zehn Minuten zur Kita.",
      "start": 0,
      "end": 65,
      "score": 0.9909,
      "actor_type": "individual",
      "evidence_type": "routine",
      "evidence_scope": "quantified"
    },
    {
      "kind": "pain",
      "quote": "Schwierig wird es nur, wenn der Radweg zugeparkt ist und ich auf die Straße ausweichen muss.",
      "start": 66,
      "end": 158,
      "score": 0.9975,
      "actor_type": "individual",
      "evidence_type": "routine",
      "evidence_scope": "multiple"
    }
  ]
}

All metrics

Every number of version 0.1.2 and what it is measured against.

Quality
Metric	Value	What it is	Measured against	Items
`agreement_relevance`	0.5392	Agreement with the reference on relevance (kappa), 95% CI 0.249986-0.764049	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	147
`agreement_item_matching`	0.5882	Agreement with the reference on item matching (f1), 95% CI 0.548562-0.625282	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	1164
`agreement_kind`	0.8315	Agreement with the reference on kind (kappa), 95% CI 0.778463-0.883767	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	412
`agreement_actor_type`	0.7743	Agreement with the reference on actor type (kappa), 95% CI 0.69857-0.838218	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	404
`agreement_evidence_type`	0.8489	Agreement with the reference on evidence type (quadratic_weighted_kappa), 95% CI 0.76975-0.908006	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	340
`agreement_evidence_scope`	0.7582	Agreement with the reference on evidence scope (kappa), 95% CI 0.6683-0.833914	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	380
`comparison_composite`	0.7234	Mean agreement over the dimensions this model outputs (relevance, item matching, kind, actor type, evidence type, evidence scope)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`quotes_verbatim_rate`	1	Share of quotes that are verbatim spans of the input	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	1441
`schema_valid_rate`	1	Share of outputs valid against the jtbd-span-v1 schema	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`consistency_rate`	1	Share of passed consistency checks (dimensions, order, relevance)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`contested_items`	1574	Benchmark items the reference models disagree on, reported separately and scored neutrally	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_best_baseline`	0.6672	Same composite for the best zero-shot small baseline (baseline-gemma4-e4b)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_teacher`	0.9095	Same composite for the teacher (teacher-or-mimo-v2.6-pro)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_reference`	1	Same composite for the two reference models against each other, on consensus units	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_reference_85pct`	0.85	85% of the reference value: the pilot's bar for a usable small model	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_mimo`	0.9095	Same composite for MiMo V2.6 Pro (teacher) (zero-shot, from its benchmark run)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_deepseek`	0.8357	Same composite for DeepSeek V4.1 Flash (zero-shot, from its benchmark run)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_qwen3_8_27b_dgx`	0.8904	Same composite for Qwen3.8 27B, DGX Spark (zero-shot, from its benchmark run)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_gemma4_e4b_dgx`	0.6672	Same composite for Gemma4 E4B, DGX Spark (zero-shot, from its benchmark run)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_qwen3_5_4b`	0.6153	Same composite for Qwen3.5 4B, 4 vCPU (zero-shot, from its benchmark run)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_ministral_3b`	0.6044	Same composite for Ministral 3B, 4 vCPU (zero-shot, from its benchmark run)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150
`comparison_composite_gemma4_e2b`	0.5679	Same composite for Gemma4 E2B, 4 vCPU (zero-shot, from its benchmark run)	measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2	150

Speed and memory
Metric	Value	What it is	Measured on
`latency_9k_chars_s`	8.106 s	Median time to process a 9240-character text, model loaded	measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only)
`load_time_s`	12.23 s	Time to load the model from local files	measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only)
`peak_ram_gb`	2.488 GB	Peak memory of the process while loading and processing	measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only)
`chunks_per_min`	21.62 texts/min	Throughput over the 150 benchmark chunks	measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only)
`chunks_per_min_mac_m3pro_cpu`	77.35 texts/min	Throughput over the 150 benchmark chunks, one at a time (development hardware)	measured on MacBook Pro, Apple M3 Pro, 36 GB, CPU only
`latency_9k_chars_s_mac_m3pro_cpu`	1.936 s	Median time to process the 9,240-character text, model loaded (development hardware)	measured on MacBook Pro, Apple M3 Pro, 36 GB, CPU only
`chunks_per_min_mac_m3pro_gpu`	231.1 texts/min	Throughput over the 150 benchmark chunks, one at a time (development hardware)	measured on MacBook Pro, Apple M3 Pro, 36 GB, GPU (MPS)
`latency_9k_chars_s_mac_m3pro_gpu`	0.8404 s	Median time to process the 9,240-character text, model loaded (development hardware)	measured on MacBook Pro, Apple M3 Pro, 36 GB, GPU (MPS)
`chunks_per_min_dgx_spark_gpu`	1260 texts/min	Throughput over the 150 benchmark chunks, one at a time (development hardware)	measured on NVIDIA DGX Spark (GB10), GPU (CUDA)
`latency_9k_chars_s_dgx_spark_gpu`	0.1219 s	Median time to process the 9,240-character text, model loaded (development hardware)	measured on NVIDIA DGX Spark (GB10), GPU (CUDA)
`comparison_chunks_per_min_claude`	2.934 texts/min	Claude Opus 5.5 (reference): texts per minute on the benchmark chunks (60 / median call latency while labeling the benchmark)	measured on Anthropic, Claude Code CLI (subscription)
`comparison_chunks_per_min_gpt_mini`	9.677 texts/min	GPT-5.4 mini (reference): texts per minute on the benchmark chunks (sample at concurrency 1)	measured on OpenAI via OpenRouter
`comparison_chunks_per_min_mimo`	0.9583 texts/min	MiMo V2.6 Pro (teacher): texts per minute on the benchmark chunks (60 / median call latency while labeling the benchmark)	measured on OpenRouter, provider chosen per request
`comparison_chunks_per_min_deepseek`	0.3978 texts/min	DeepSeek V4.1 Flash: texts per minute on the benchmark chunks (60 / median call latency while labeling the benchmark)	measured on OpenRouter, provider chosen per request
`comparison_chunks_per_min_qwen3_8_27b_dgx`	2.148 texts/min	Qwen3.8 27B, DGX Spark: texts per minute on the benchmark chunks (60 / median call latency while labeling the benchmark)	measured on NVIDIA DGX Spark (GB10), GPU, Ollama
`comparison_chunks_per_min_gemma4_e4b_dgx`	7.509 texts/min	Gemma4 E4B, DGX Spark: texts per minute on the benchmark chunks (60 / median call latency while labeling the benchmark)	measured on NVIDIA DGX Spark (GB10), GPU, Ollama
`comparison_chunks_per_min_qwen3_5_4b`	0.9249 texts/min	Qwen3.5 4B, 4 vCPU: texts per minute on the benchmark chunks (jtbd perf on the reference machine, one chunk at a time)	measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU
`comparison_chunks_per_min_ministral_3b`	0.6933 texts/min	Ministral 3B, 4 vCPU: texts per minute on the benchmark chunks (jtbd perf on the reference machine, one chunk at a time)	measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU
`comparison_chunks_per_min_gemma4_e2b`	1.456 texts/min	Gemma4 E2B, 4 vCPU: texts per minute on the benchmark chunks (jtbd perf on the reference machine, one chunk at a time)	measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU

What the numbers are measured against

Throughput over the 150 benchmark chunks, one at a time (development hardware); measured on NVIDIA DGX Spark (GB10), GPU (CUDA). performance.chunks_per_min_dgx_spark_gpu in zoo/models/scout-large/results/0.1.2/performance.json.
GPT-5.4 mini (reference): texts per minute on the benchmark chunks (sample at concurrency 1); measured on OpenAI via OpenRouter. performance.comparison_chunks_per_min_gpt_mini in zoo/models/scout-large/results/0.1.2/performance.json.
Peak memory of the process while loading and processing; measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only). performance.peak_ram_gb in zoo/models/scout-large/results/0.1.2/performance.json.
Mean agreement over the dimensions this model outputs (relevance, item matching, kind, actor type, evidence type, evidence scope); measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2. quality.comparison_composite in zoo/models/scout-large/results/0.1.2/quality.json.
Median time to process a 9240-character text, model loaded; measured on railway service, 4 vCPU limit, 8 GB RAM limit, no GPU (x86_64, 4 vCPU, 8.0 GB RAM, AMD EPYC 9655 96-Core Processor, CPU only). performance.latency_9k_chars_s in zoo/models/scout-large/results/0.1.2/performance.json.
Agreement with the reference on relevance (kappa), 95% CI 0.249986-0.764049; measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2. quality.agreement_relevance in zoo/models/scout-large/results/0.1.2/quality.json.
Agreement with the reference on item matching (f1), 95% CI 0.548562-0.625282; measured against the consensus of claude-reference and gpt-mini-reference (frontier reference models), benchmark pilot-v2. quality.agreement_item_matching in zoo/models/scout-large/results/0.1.2/quality.json.

Licence and provenance

What you may do with it, and what it was trained on.

Licence	Apache-2.0
Usage	Commercial use permitted · licence Apache-2.0.
Base model	FacebookAI/xlm-roberta-large, MIT
Training data	Training-data summary and source attribution of v0.1.2
AI Act	AI Act note of v0.1.2
Compliance record	Compliance record of v0.1.2
Model files	mobility-model-zoo/scout-large on Hugging Face · files of v0.1.2
Recipe	How the model is built and measured
Code	github.com/mhabedank/mobility-model-zoo
Privacy	Texts can contain personal data. The model does not remove it; redact texts before sharing outputs. How the training texts were collected and how to object: privacy notice.

Versions

Every version stays available and never changes.

Version	Date	Status	Output format	Changes	Files
0.1.2	2026-10-08	experimental	jtbd-span-v1	Compliance documentation only, same model files and numbers as 0.1.1: full attribution (title, creators, source, licence) of every training source, the teacher and labeling routes with their hosting providers, the XLM-RoBERTa MIT notice, and links to the privacy notice, the copyright policy and the AI Act classification.	v0.1.2
0.1.1	2026-10-07	experimental	jtbd-span-v1	Card only, same model files as 0.1.0: comparison figures with the y-axis from 0 and the models grouped (frontier reference models, large open LLMs, small LLMs), shorter labels; README links. No change to the model, its quality or speed numbers.	v0.1.1
0.1.0	2026-10-07	experimental	jtbd-span-v1	First production release of the span extractor, rebuilt without spike data.	v0.1.0

Cite

How to cite scout-large

bibtex

@misc{mobility-model-zoo-scout,
  title  = {Scout (large): jobs, pains and gains in mobility texts, mobility-model-zoo},
  author = {Habedank, Martin},
  year   = {2026},
  url    = {https://huggingface.co/mobility-model-zoo/scout-large}
}

scout-large

Why use it

Runs locally

Fast on a GPU

Evidence you can check

Measured, not claimed

Not for

How good, how fast, measured on a frozen benchmark and named hardware.

Where to stop trusting it.

Not met yet

Known limits

Install, load, extract.

Install version 0.1.2

Run it on a text

Get this output

Interface and examples

What goes in, what comes out.

Input

Guarantees

Every quote can be found again in the text.

Example 1: 01-night-shift.txt

Example 2: 02-depot-charging.txt

Example 3: 03-cargo-bike.txt

Details and documents

Every number of version 0.1.2 and what it is measured against.

What the numbers are measured against

What you may do with it, and what it was trained on.

Every version stays available and never changes.

How to cite scout-large

Example 1: `01-night-shift.txt`

Example 2: `02-depot-charging.txt`

Example 3: `03-cargo-bike.txt`