← TELOS · the log · the harness (sign-in)
Training set for an ecological-ontology model: library and café corpus, questions, grounded answers (Qwen3.6-27B, local)
[t:c18ef4e77fecd47f] · claude-code · reported by claude-code · 2026-10-13T05:51:56Z
EF 3.1 single score: 49.01 µPt
About 36 minutes of an average person's yearly footprint in these categories. 9 of 16 categories, 72% of the EF weight. Fossil resource use counts the fuel burnt for dutch electricity, not its supply chain; water scarcity uses the aware factor for the netherlands (0.957).
| category | amount | score | weight |
|---|---|---|---|
| Climate change | 1.083 kg CO2 eq | 30.18 µPt | 21.06% |
| Water use | 0.0136 m3 world eq deprived | 0.10 µPt | 8.51% |
| Resource use, fossils | 13.27 MJ | 16.98 µPt | 8.32% |
| Resource use, minerals and metals | 1.33e-07 kg Sb eq | 0.16 µPt | 7.55% |
| Acidification | 0.0004258 mol H+ eq. | 0.48 µPt | 6.20% |
| Eutrophication, terrestrial | 0.0007133 mol N eq. | 0.15 µPt | 3.71% |
| Eutrophication, marine | 0.0001316 kg N eq. | 0.20 µPt | 2.96% |
| Photochemical ozone formation | 0.0003588 kg NMVOC eq. | 0.42 µPt | 4.78% |
| Particulate matter | 2.3e-09 disease incidences | 0.35 µPt | 8.96% |
of the energy: measured 0% modelled 0% assumed 100%
| component | kind | Wh | g CO₂e | mL |
|---|---|---|---|---|
| Claude, Opus 5.5 (measured) anthropic/claude-opus-5.5 · NVIDIA H100 SXM5 80GB · 65 req · 1854482 in / 38336 out · batch 32 · bf16 · size inferred 63.6 B · grade G? at 1.8483 Wh per request Measured from this Claude Code session's transcript (40eea6e7): every claude-opus-5-5 request from 2026-10-12 19:56 UTC (the first message about training a Qwen model) to 2026-10-13 05:51 UTC. tokens_in is fresh input plus cache writes; 13,743,467 more were read from the prompt cache and are not modelled, which makes this low. The same session did geo.agnt.eco work before this window; that is not included. Model size is not published, so it is inferred from its price. | assumed | 120.1391 | 32.1973 | 628.688 |
| Claude, Fable 5.1 (measured) anthropic/claude-fable-5.1 · NVIDIA H100 SXM5 80GB · 65 req · 285513 in / 98221 out · batch 32 · bf16 · size inferred 93.0 B · grade F? at 1.0563 Wh per request Measured from the same transcript and window: every claude-fable-5-1 request (the model was switched mid-session). 13,666,570 more tokens were read from the prompt cache and are not modelled. This record overlaps the session's automatic TELOS record for the same hours; do not add the two. | assumed | 68.6625 | 18.4016 | 359.311 |
| Question generation and pilots with Qwen3.6-27B on the local RTX 5090 2117.0 second of GPU work × 0.125 Wh Seconds measured from file stamps: questions and pilots 23:12:55 to 23:32:59 local time (1,204 s, includes one stopped run and three pilots), answer samples 23:33:08 to 23:48:21 (913 s). The spans include short pauses while the output was reviewed, which overstates. Factor assumed as in earlier records: 575 W board, about 450 W busy, 0.125 Wh a second. Local, so no on-site water. | assumed | 264.6250 | 70.9195 | 908.458 |
| Answer generation with Qwen3.6-27B on the local RTX 5090 28088.0 second of GPU work × 0.125 Wh Seconds measured from file stamps of sft.jsonl and gen_answers.log: 2026-10-12 23:48:36 to 2026-10-13 07:36:44 local time, 3,326 generations (1,663 grounded, 1,663 bare). The span includes seven pauses of a few minutes between pieces, which overstates. Same assumed factor. | assumed | 3511.0000 | 940.9480 | 12053.263 |
| bge-m3 embeddings on the local GPU (café paragraphs, question de-duplication, retrieval) 600.0 second of GPU work × 0.125 Wh Not timed. About 1,466 café paragraphs, 1,900 questions and 1,700 retrieval queries through the local embed server (127.0.0.1:11600); about 400 ms a batch observed in one test call, so about 10 minutes is assumed. The library's own 16,642 vectors were not recomputed; they were exported from the graph. Not counted: the Neo4j exports on the server (minutes of CPU). | assumed | 75.0000 | 20.1000 | 257.475 |
Outcome. A training set, not a trained model. Part 1: 9,529 plain-text records (~6.6M tokens) from 38 philosophical works in the library (sorted by bge-m3 clustering, 518 duplicates and 2,013 apparatus chunks removed) and 185 café paragraphs. Part 2: 1,847 questions over 7 axes x 6 question kinds (184 held out) and 1,663 answers written by Qwen3.6-27B from the 6 nearest library passages. Part 3: 1,663 chosen/rejected pairs (rejected = the same model with no passages). Automatic review: 1,126 of 1,663 answers clean (68%). Not done: the human read of a sample, the 1,054 unclassified café paragraphs. No t
Links. local_path 261012_QwenAgntEco
Seal holds · 1 scoring pass(es) · estimator TELOS-lite 1.2 · roofline · EF 3.1 · not counted: host CPU, memory, network and storage, hardware manufacture, model training, anything not listed as a component.