← TELOS · the log · the harness (sign-in)
QLoRA training of Qwen3.6-27B on the ecological-ontology training set, parts 1 and 2, local RTX 5090
[t:717b8cdcfb5d877c] · claude-code · reported by claude-code · 2026-10-13T16:39:58Z
· superseded by 110d31d7c0b4dfad
EF 3.1 single score: 20.83 µPt
About 15 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 | 0.46 kg CO2 eq | 12.83 µPt | 21.06% |
| Water use | 0.005794 m3 world eq deprived | 0.04 µPt | 8.51% |
| Resource use, fossils | 5.637 MJ | 7.21 µPt | 8.32% |
| Resource use, minerals and metals | 5.652e-08 kg Sb eq | 0.07 µPt | 7.55% |
| Acidification | 0.0001809 mol H+ eq. | 0.20 µPt | 6.20% |
| Eutrophication, terrestrial | 0.0003031 mol N eq. | 0.06 µPt | 3.71% |
| Eutrophication, marine | 5.59e-05 kg N eq. | 0.08 µPt | 2.96% |
| Photochemical ozone formation | 0.0001524 kg NMVOC eq. | 0.18 µPt | 4.78% |
| Particulate matter | 9.773e-10 disease incidences | 0.15 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 · 108 req · 1201711 in / 62426 out · batch 32 · bf16 · size inferred 63.6 B · grade F? at 0.8280 Wh per request Measured from this Claude Code session's transcript (40eea6e7): every claude-opus-5-5 request from 2026-10-13 05:51 UTC (after the training-set record) to 16:39 UTC. tokens_in is fresh input plus cache writes; 41,137,387 more were read from the prompt cache and are not modelled, which makes this low. The window also holds the TELOS report for the training set and a patch to the TELOS receipt page. This record overlaps the session's automatic TELOS record; do not add the two. | assumed | 89.4230 | 23.9654 | 467.951 |
| Part 1 training: plain text, 336 steps, QLoRA on the local RTX 5090 8050.0 second of GPU work × 0.125 Wh 7,904 s measured by the trainer (run_1791897130.json), plus about 150 s for loading the model three times (two starts stopped on a data-loader error before training). 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 | 1006.2500 | 269.6750 | 3454.456 |
| Part 2 training: question-answer pairs, 416 steps, QLoRA on the local RTX 5090 3696.0 second of GPU work × 0.125 Wh Span of train_sft.log, 15:25:09 to 16:26:45 local time: 3,573 s of training measured by the trainer plus a wrong start of a few minutes (it resumed the test checkpoint and was stopped) and model loading. Same assumed factor. | assumed | 462.0000 | 123.8160 | 1586.046 |
| Test run and demonstrations on the local RTX 5090 1269.0 second of GPU work × 0.125 Wh 30-step test run: 405 s (log span 12:07:54 to 12:14:39). Demonstration of 6 held-out questions, trained and untrained: 564 s (log span 17:57:46 to 18:07:10). The two single questions (name, place and thought): about 300 s, estimated from file times. Same assumed factor. Not counted: the 52 GB model download (network transfer, no factor available). | assumed | 158.6250 | 42.5115 | 544.560 |
Outcome. A trained LoRA adapter (233M trainable parameters, 0.86% of the model) on Qwen3.6-27B in 4-bit: part 1, 336 steps over 8,560 plain-text records from the philosophical library (loss 3.40 to about 2.55); part 2, 416 steps over 1,663 question-answer pairs, two passes (loss 2.05 to about 1.36). Tested on 6 held-out questions against the untrained model, plus two single questions (a self-chosen name; where place and thought differ). Part 3 (preference pairs) is not done. The training set it was built from is the record c18ef4e77fecd47f.
Links. local_path 261012_QwenAgntEco · telos_session c18ef4e77fecd47f
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.