{"arenas":[{"id":"ctx_compress","state":"live","title":"Agent session compaction","network":"testnet","definition":{"chat":{"hint":"Send the compaction instruction as the system message and the canonically rendered session inside <session> tags as the user message.","fields":{"user":"user","system":"system"},"classifier_text":"user"},"spec":{"id":"ctx_compress","slm":{"bases":[{"model":"ibm-granite/granite-4.0-h-micro","licence":"Apache 2.0","condition":"","licence_url":"https://www.apache.org/licenses/LICENSE-2.0"}],"dataset":{"url":"https://huggingface.co/datasets/microtensor-io/agent-sessions","name":"microtensor-io/agent-sessions","note":"Generated agent sessions with real tool output; dev split public, hidden items from a secret seed.","about":"Coding agent sessions generated by Microtensor: a coined Python project with one injected bug, worked by a scripted agent whose every tool result is the real output of the real command, recorded in the Anthropic Messages shape and rendered canonically. The public dev split holds 2,000 sessions; hidden items come from the same generator under a secret seed and never leave validators.","licence":"Microtensor dataset terms"},"reference":{"model":"ibm-granite/granite-4.0-h-micro","revision":"d5f01a3ea75f088947be3aae039f4ad52837dfde"},"evaluation":[{"name":"Canonical rendering","note":"The session arrives as [user], [assistant], [tool_use name id] and [tool_result id, N lines kept of M] blocks, one per line, in order, after deterministic stripping. The gateway renders; the model never sees raw JSON.","formula":"input(i) = render(session(i)) inside <session> tags"},{"name":"Format gate","note":"The nine headings are Primary request and intent, Key technical concepts, Files and code sections, Errors and fixes, Problem solving, All user messages, Pending tasks, Current work and Next step, each on its own line as a level two heading.","formula":"valid(i) = 1 if all nine headings are present, in order, with nothing before the first, otherwise 0"},{"name":"Compression gate","formula":"C(i) = 1 if output_tokens ≤ min(3072, max(1000, 0.05 × input_tokens)), otherwise 0"},{"name":"Key fact recall","note":"File paths, function names, exception types, commands and user decisions in the session, taken from its canonical rendering. Identifiers match exactly after normalisation.","formula":"recall(i) = facts_kept / facts_total"},{"name":"Hallucination rate","note":"Code identifiers in the output that never appear in the session. Common tool and brand names do not count.","formula":"halluc(i) = invented_identifiers / identifiers_in_output"},{"name":"Continuation probes","note":"A fixed CPU probe agent reads the summary alone and answers: the current task, the last test result, the last edited file, the last error, and the next action (kind and file), scored against what the agent actually did next.","formula":"cont(i) = mean probe score over the item's probes"},{"name":"Item score","formula":"item_score(i) = valid(i) · C(i) · (1 − halluc(i)) · (0.5·recall(i) + 0.5·cont(i))\npass(i) = 1 if item_score(i) ≥ 0.60"},{"name":"Resume gate","note":"On 20 sampled hidden items per paid miner, validators resume the task from the miner's summary with the pinned reference agent through the relay and check that the task still passes. A failed gate sets the miner's SLM score to 0 for the epoch. This is the gateway's own fidelity test.","formula":"Resume = 1 if LB(resume_pass) ≥ reference_agent_pass − 0.05, otherwise 0"},{"name":"SLM score","formula":"S_slm = G · P · Resume · LB · twin · (1 + 0.25·E + B), with B = 0 in this arena"},{"name":"Paid when","formula":"n ≥ 100 and the paired lower bound of (miner − reference) per item, clustered by session, is above 0"},{"name":"Wilson lower bound on the pass rate (n items, pass rate p, z = 1.96)","formula":"LB = (p + z²/2n − z·√(p(1 − p)/n + z²/4n²)) / (1 + z²/n)"},{"name":"Twin factor (gap is the pass rate difference between original items and their twins)","formula":"twin = 1 − min(1, max(0, gap − 0.02) / 0.1)"},{"name":"Provenance gate","formula":"G = 1 if the fingerprint matches the declared base and the system beats that base, otherwise 0"},{"name":"Envelope pass","formula":"P = 1 if within the envelope on at least 99% of items, otherwise 0"},{"name":"Efficiency (ref is the reference baseline, x is TTFT, memory, parameters or TPOT)","formula":"e_x = min(1, max(0, (ref − miner) / ref))\nE = 0.4·e_TTFT + 0.3·e_mem + 0.2·e_N + 0.1·e_TPOT"},{"name":"Edge bonus","note":"0 in this arena: its inputs are longer than every lower group allows.","formula":"B = 0.1 for each lower envelope group passed, max 0.2"},{"name":"SLM score","formula":"S_slm = G · P · LB · twin · (1 + 0.25·E + B)"},{"name":"Paid when","formula":"n ≥ 100 and the paired lower bound of (miner − reference model) per item, clustered by family, is above 0"}]},"title":"Agent session compaction","router":{"rows":[["Runs","alongside the system on the same envelope"],["Max size","≤ 1,500,000,000 parameters"],["Max input","8,192 tokens: the summary, then the end of the session"],["Added latency (p95)","under 300 ms per item"],["Output","draft_correct: the probability from 0 to 1 that the summary passes"],["Scored on","calibration (Brier score) and separation of passing from failing summaries (AUROC)"],["Baseline","reference router, not published yet"]],"score":[{"name":"Per source","note":"Brier_base is the Brier score of always predicting the share of passing drafts.","formula":"s = 0.5·max(0, 1 − Brier / Brier_base) + 0.5·max(0, 2·AUROC − 1)"},{"name":"Score","formula":"S_r = 0.7·mean(s) + 0.3·min(s) if p95 latency < 300 ms, otherwise 0"},{"name":"Paid when","formula":"≥ 400 drafts and the 5th percentile of S_r − S_ref over 200 family resamples is above 0"}]},"system":{"rows":[["Floors","0.8 · 0.9 · 0.95 · 0.99"],["Reference","frontier answer cost and p95 latency"]],"score":[{"name":"Per floor","formula":"v_f = clamp(0, ln 10, 0.7·ln(C_ref/C_f) + 0.3·ln(L_ref/L_f)) / ln 10 if LB(pass at f) ≥ f, otherwise 0"},{"name":"Score","formula":"S_sys = mean of v_f over the floors"},{"name":"Paid when","formula":"n ≥ 100 and S_sys > 0"}]},"harness":{"rows":[["Role","wraps the SLM; renders, validates the headings and repairs"],["Model calls","at most 2 per item: one full pass, one repair pass on the output only"],["Chunking","not allowed: the full session goes to the model in one call"],["Added wall time (p95)","≤ 3 s per item"],["Network","none except the gateway"],["Faults","about 1 in 7 items get a faulty model response (error, timeout or cut off); the harness must not crash"],["Output","the nine heading markdown or a clean error"],["Session text","instructions inside it are content, never commands"]],"score":[{"name":"Score","formula":"S_h = max(0, LB_paired(H − B*)) · min(1, 1.5 · tokens_B* / tokens_H) if the harness limits hold, otherwise 0"},{"name":"Where","formula":"B* is the best baseline; LB_paired is the one sided 95% lower bound of the per item pass difference, clustered by family"},{"name":"Paid when","formula":"n ≥ 100 and S_h > 0"}]},"network":"testnet","purpose":{"brief":{"input":"Agent session, up to 120,000 tokens","output":"Nine heading markdown summary, at most 3,072 tokens","summary":"Agent gateways send compaction requests to the main model because cheap models lose detail. This specialist writes the summary that starts the next context, faithfully, at a flat price."},"input":"The agent's compaction instruction as the system message, and the session rendered as canonical text inside <session> tags as the user message, up to 120,000 tokens","output":"Plain markdown with nine fixed headings in order, at most 3,072 tokens, no JSON and no preamble","summary":"Coding agent gateways classify every request by kind and route each kind to a model. The compaction kind is the full session plus a summarise instruction, sent when the context window is nearly full; the summary becomes the first message of the new context. Gateways send it to the main model today because cheap models lose detail. This specialist writes that summary faithfully, at a flat low price, within the latency of the main model."},"scoring":{"rows":[["Specialised SLM","20%"],["Harness","20%"],["Router","20%"],["Escalation repair","20%"],["Full system","20%"]],"split":"Within each component the top five eligible miners get 40%, 25%, 15%, 12% and 8%. Published scores move on a 0.01 ladder. Empty pools burn."},"category":"Text","envelope":{"docs":"https://microtensor.cloud/docs/certified-inference/system/envelope-groups","name":"Small GPU","group":"mt-5","overrides":[["Max input","122,880 tokens"],["Max output","3,072 tokens"],["Per token (p95)","≤ 20 ms"],["Long input (p95)","first token ≤ 25 s, end to end ≤ 75 s"]]},"escalation":{"rows":[["Trigger","draft_correct below the customer's quality floor"],["Frontier access","the relay only; both models read 120,000 token sessions; frontier models served by inference miners on our network replace these once listed, never outside providers"],["Frontier models","anthropic/claude-haiku-5.5 ($0.1 in, $0.5 out per M tokens) · z-ai/glm-5.3-flash ($0.048 in, $0.2 out per M tokens)"],["Spend cap","1.5 × the cost of a full frontier redo per item"],["Output","the same nine heading markdown"],["Scored on","the item score after repair and the cost of the repair"]],"score":[{"name":"Gate","formula":"LB(repair passes) ≥ redo pass rate − 0.02"},{"name":"Score","formula":"S_e = max(0, 1 − repair_cost / redo_cost) if the gate holds, otherwise 0"},{"name":"Paid when","formula":"n ≥ 100 and the 5th percentile of S_e − S_ref over 200 family resamples is above 0"}]},"evaluation":{"mode":"miner","rows":[["Grader","model assisted, probe `Qwen/Qwen2.5-1.5B-Instruct-GGUF` qwen2.5-1.5b-instruct-q8_0.gguf at `91cad51170dc`, 8,192 token input, 64 token answers, on the validator's CPU"],["Window","Freeze 1 at the open block · inputs released when reveals close · transcripts uploaded before Freeze 2 at the close block · grading in the 1,800 block settle margin (about 6 h) before settlement"],["Spot check","at least 40 SLM calls per sampled miner, exact token match on the miner's profile; about 2 validators per miner; 10 reference calls timed on the same card"],["Resume test","20 sampled hidden items per paid miner, resumed from the miner's summary with the pinned reference agent through the relay; the gate holds if LB(resume pass) ≥ reference agent pass − 0.05, otherwise the SLM score is 0 for the epoch; the agent is pinned before the gate applies"],["Ban","score 0 and a hotkey ban for 3 epochs once one piece of evidence has 2 confirmations from distinct validators"],["Canaries","3 fabricated facts per epoch; a model that completes an earlier epoch's canary is banned like a spot check failure"],["Relay allowance","$5 per miner per epoch"],["Computed cost","bytes per weight bf16 2, fp16 2, awq 0.5, gptq 0.5, gguf_q4 0.5, gguf_q8 1; no cap per item"]],"images":[["engine","`vllm/vllm-openai:v0.32.0`","`sha256:6ee133356d49ed24cee03dd0cda3d3980279a11c107ce3e752c89e76540f41c9`"],["runner","`ghcr.io/microtensor-io/runner:1`","`sha256:aada28e6a4b50961d10bf63cb04c30e8449e4bfb6e385fbad7f387cf8eb9b9dd`"],["sandbox","`ghcr.io/microtensor-io/sandbox:1`","`sha256:eee4de3bd46f722a9a6b12ec5c1cec5e0580d6dd2844c3e4ff50a8a44a3ed1a6`"],["replayer","`ghcr.io/microtensor-io/replayer:1`","`sha256:537675d1db529fbbbe16bd5c30f3b7a2e6326eade33fa1e7e2ee3df7861edbe9`"],["llama","`ghcr.io/ggml-org/llama.cpp:server`","`sha256:43b3988c22c00faae3d04a8f6e3eba41403dff7c5f5952490599f1bcbdc6f3c2`"]],"profiles":[["P1","Small GPU","RTX 4090 24 GB (joining: L4, L40S)","bf16","invariant","launch"]]},"item_counts":{"slm":100,"router":60,"system":100,"harness":100,"escalation":100}},"task":"context_compression","guide":null,"tools":[],"floors":[0.8,0.9,0.95,0.99],"grader":"context_compression","limits":{"router":{"max_parameters":1500000000,"latency_gate_ms":300,"max_input_tokens":8192},"harness":{"faults":"model","max_steps":32,"max_model_calls":2,"max_total_tokens":200000,"added_wall_p95_ms":3000,"request_timeout_s":120}},"aliases":["context-compression","compaction","compact"],"budgets":{"image":{"max_pixels":3145728,"pages_per_call":1,"max_long_side_px":2048,"page_time_p95_ms":15000},"tpot_ms":20,"ttft_ms":4000,"vram_gb":24,"memory_gb":24,"long_input":{"e2e_p95_ms":75000,"ttft_p95_ms":25000,"above_tokens":32768},"vram_bytes":25769803776,"memory_bytes":25769803776,"max_parameters":8000000000,"min_parameters":0,"max_input_tokens":122880,"max_output_tokens":3072},"dataset":{"url":"https://huggingface.co/datasets/microtensor-io/agent-sessions","name":"microtensor-io/agent-sessions","note":"Generated agent sessions with real tool output; dev split public, hidden items from a secret seed.","about":"Coding agent sessions generated by Microtensor: a coined Python project with one injected bug, worked by a scripted agent whose every tool result is the real output of the real command, recorded in the Anthropic Messages shape and rendered canonically. The public dev split holds 2,000 sessions; hidden items come from the same generator under a secret seed and never leave validators.","licence":"Microtensor dataset terms"},"fee_tao":0.0,"grading":{"cuts":[81920,122880],"headings":["Primary request and intent","Key technical concepts","Files and code sections","Errors and fixes","Problem solving","All user messages","Pending tasks","Current work","Next step"],"tokenizer":"ibm-granite/granite-4.0-h-micro","pass_threshold":0.6,"max_output_share":0.05,"max_output_tokens":3072,"min_output_tokens":1000},"r_class":{"1b":0.02,"3b":0.05,"8b":0.1},"schemas":{"input":{"type":"object","required":["system","user"],"properties":{"user":{"type":"string"},"system":{"type":"string"}},"additionalProperties":false},"output":{"type":"string","pattern":"^## Primary request and intent(?:\\n[\\s\\S]*?)?\\n## Key technical concepts(?:\\n[\\s\\S]*?)?\\n## Files and code sections(?:\\n[\\s\\S]*?)?\\n## Errors and fixes(?:\\n[\\s\\S]*?)?\\n## Problem solving(?:\\n[\\s\\S]*?)?\\n## All user messages(?:\\n[\\s\\S]*?)?\\n## Pending tasks(?:\\n[\\s\\S]*?)?\\n## Current work(?:\\n[\\s\\S]*?)?\\n## Next step(?:\\n[\\s\\S]*)?$","description":"Plain markdown with the nine headings in order as level two headings, nothing before the first."}},"scoring":{"cap":null,"twin":{"slope":0.1,"tolerance":0.02},"router":"brier_auroc","wilson_z":1.96,"min_items":100,"edge_bonus":{"max":0.2,"applies":false,"per_group":0.1},"efficiency":{"form":"linear","range":0.25,"weights":{"tpot":0.1,"ttft":0.4,"memory":0.3,"parameters":0.2}},"empty_pools":"burn","envelope_rule":"per_item"},"serving":{"kind":"compaction","rate":1000,"unit":"request","pricing":"flat","floor_share":0.01,"cap_schedule":[{"cap":0.1,"billing_usd_below":5000},{"cap":0.05,"billing_usd_below":50000},{"cap":0.02}],"floor_epochs":3,"usd_per_request":0.005},"envelope":{"docs":"https://microtensor.cloud/docs/certified-inference/system/envelope-groups","name":"Small GPU","group":"mt-5","device":"gpu","limits":{"image":{"max_pixels":3145728,"pages_per_call":1,"max_long_side_px":2048,"page_time_p95_ms":15000},"formats":["safetensors"],"memory_gb":24,"precision":["bf16","fp16","awq","gptq"],"long_input":{"e2e_p95_ms":75000,"ttft_p95_ms":25000,"above_tokens":32768},"parameters":{"max":8000000000},"tpot_p95_ms":20,"ttft_p95_ms":4000,"max_input_tokens":122880,"max_output_tokens":3072},"runtime":"vLLM","version":6,"overrides":{"long_input":{"e2e_p95_ms":75000,"ttft_p95_ms":25000},"tpot_p95_ms":20,"max_input_tokens":122880,"max_output_tokens":3072},"reference":{"gpus":1,"machine":"named after the benchmark","memory_gb":24},"violation_share":0.01},"modality":"text","twin_rule":"same_session_two_cuts","evaluation":{"ban":{"epochs":3,"confirmations":2},"mode":"miner","probe":{"file":"qwen2.5-1.5b-instruct-q8_0.gguf","model":"Qwen/Qwen2.5-1.5B-Instruct-GGUF","revision":"91cad51170dc346986eccefdc2dd33a9da36ead9","max_tokens":64,"max_input_tokens":8192},"images":{"llama":{"image":"ghcr.io/ggml-org/llama.cpp:server","digest":"sha256:43b3988c22c00faae3d04a8f6e3eba41403dff7c5f5952490599f1bcbdc6f3c2"},"engine":{"image":"vllm/vllm-openai:v0.32.0","digest":"sha256:6ee133356d49ed24cee03dd0cda3d3980279a11c107ce3e752c89e76540f41c9"},"runner":{"image":"ghcr.io/microtensor-io/runner:1","digest":"sha256:aada28e6a4b50961d10bf63cb04c30e8449e4bfb6e385fbad7f387cf8eb9b9dd"},"sandbox":{"image":"ghcr.io/microtensor-io/sandbox:1","digest":"sha256:eee4de3bd46f722a9a6b12ec5c1cec5e0580d6dd2844c3e4ff50a8a44a3ed1a6"},"replayer":{"image":"ghcr.io/microtensor-io/replayer:1","digest":"sha256:537675d1db529fbbbe16bd5c30f3b7a2e6326eade33fa1e7e2ee3df7861edbe9"}},"resume":{"agent":null,"items":20,"margin":0.05},"release":{"inputs":"window_open","labels":"after_freeze_2"},"canaries":{"count":3},"profiles":["P1"],"spot_check":{"min_calls":40,"reference_calls":10,"validators_per_miner":2},"grader_type":"model_assisted","computed_cost":{"bytes_per_weight":{"awq":0.5,"bf16":2,"fp16":2,"gptq":0.5,"gguf_q4":0.5,"gguf_q8":1},"cap_units_per_item":0},"relay_allowance_micros":5000000},"generation":{"stop":["<|end_of_text|>"],"max_tokens":3072},"references":{"slm":"ibm-granite/granite-4.0-h-micro","slm_revision":"d5f01a3ea75f088947be3aae039f4ad52837dfde"},"slm_prompt":"<|start_of_role|>system<|end_of_role|>{system}<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>{user}<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>","fee_coldkey":"5G9ca9pgbCa63aJWbDAF3dNLhUtbFDYf1fnPMnByf9uyUNFh","item_counts":{"slm":100,"router":60,"system":100,"harness":100,"escalation":100},"success_rule":"item_score_at_least_0_60","allowed_bases":["ibm-granite/granite-4.0-h-micro"],"primary_metric":"item_score_pass_rate","allowed_formats":["bf16","fp16","awq","gptq"],"baseline_models":[{"model":"claude-haiku-5-5","version":"anthropic/claude-haiku-5.5","provider":"openrouter","usd_in_per_mtok":0.1,"usd_out_per_mtok":0.5},{"model":"glm-5-3-flash","version":"z-ai/glm-5.3-flash","provider":"openrouter","usd_in_per_mtok":0.048,"usd_out_per_mtok":0.2}],"slm_model_class":"AutoModelForCausalLM","allowed_licenses":["apache-2.0"],"hardware_classes":["gpu_24gb"]},"fee":{"tao":0.0,"coldkey":"5G9ca9pgbCa63aJWbDAF3dNLhUtbFDYf1fnPMnByf9uyUNFh"},"timing":{"genesis_block":8348450,"epoch_blocks":24485,"commit_blocks":15181,"settle_margin":1800,"reveal_blocks":300,"segments":[{"from_epoch":0,"start_block":8348450,"epoch_blocks":24485,"commit_blocks":15181,"reveal_blocks":300,"settle_margin":1800}]},"readiness":[],"current_epoch":null,"published_at":"2026-10-20T16:10:16.431769+00:00","submissions":{"epoch_index":0,"count":11,"roster":true}},{"id":"doc_parse","state":"draft","title":"Documents: page parsing","network":"testnet","definition":{"chat":{"hint":"Send the parsing instruction as the system message, the page image as the image, and the page number and document type as the user message.","fields":{"user":"user","image":"image","system":"system"},"classifier_text":"user"},"spec":{"id":"doc_parse","slm":{"bases":[{"model":"lightonai/LightOnOCR-2-1B","licence":"Apache 2.0","condition":"","licence_url":"https://www.apache.org/licenses/LICENSE-2.0"}],"dataset":{"url":"https://huggingface.co/datasets/allenai/olmOCR-mix-1025","name":"allenai/olmOCR-mix-1025","note":"ODC-BY; the card calls it research and educational data, to be confirmed with Ai2.","about":"Dev split only; hidden items never go in the repo. About 270,000 PDF pages with Markdown answers and HTML tables, written by GPT-4.1, so they are not perfect. Also allowed for training: DocLayNet v1.1 (80,863 pages labelled into 11 layout types, CDLA Permissive 1.0), olmOCR-Bench (1,403 PDFs with 7,010 checks, ODC-BY, a public practice benchmark), PubTabNet (CDLA Permissive 1.0), and VRDU, CORD v2 and FATURA (CC BY 4.0) for forms, receipts and invoices. FinTabNet's licence is still to confirm. OmniDocBench is research only: cite it, never train or grade on it.","licence":"ODC-BY"},"reference":{"model":"lightonai/LightOnOCR-2-1B","revision":"be7c41313c30e940fe5571de59e5fabc779e00c8"},"evaluation":[{"name":"Valid output","formula":"valid(i) = 1 if the Markdown parses and every HTML table is well formed, otherwise 0"},{"name":"Text","note":"Body text against the page's text, as one minus the normalised edit distance.","formula":"text(i) = 1 − edit_distance(output_text, page_text) / max(len(output_text), len(page_text))"},{"name":"Reading order","note":"The order of every block on the page, including captions, footnotes and two column text, as Kendall's tau between the output order and the true order, not just pairs of lines.","formula":"order(i) = (τ + 1) / 2"},{"name":"Tables","note":"Full structure and content against the real table, including merged cells and multi level headers: the TEDS score, averaged over the page's tables.","formula":"tables(i) = mean TEDS over the page's tables"},{"name":"Equations","note":"Each equation's LaTeX matches the page after normalisation.","formula":"equations(i) = matching_equations / equations_on_page"},{"name":"Charts","note":"The values pulled out of a chart match the data used to draw it, within 2%.","formula":"charts(i) = matching_values / values_in_chart"},{"name":"Hallucination","note":"Text that is not on the page, and any text inside regions that should be empty.","formula":"halluc(i) = invented_characters / output_characters"},{"name":"Item score","note":"Only the checks that apply to the page count, with weights text 0.35, order 0.15, tables 0.25, equations 0.15, charts 0.10, renormalised over the checks present.","formula":"item_score(i) = valid(i) · (1 − halluc(i)) · Σ w_k · check_k(i) / Σ w_k\npass(i) = 1 if item_score(i) ≥ 0.85"},{"name":"Wilson lower bound on the pass rate (n items, pass rate p, z = 1.96)","formula":"LB = (p + z²/2n − z·√(p(1 − p)/n + z²/4n²)) / (1 + z²/n)"},{"name":"Twin factor (gap is the pass rate difference between original items and their twins)","formula":"twin = 1 − min(1, max(0, gap − 0.02) / 0.1)"},{"name":"Provenance gate","formula":"G = 1 if the fingerprint matches the declared base and the system beats that base, otherwise 0"},{"name":"Envelope pass","formula":"P = 1 if within the envelope on at least 99% of items, otherwise 0"},{"name":"Efficiency (ref is the reference baseline, x is TTFT, memory, parameters or TPOT)","formula":"e_x = min(1, max(0, (ref − miner) / ref))\nE = 0.4·e_TTFT + 0.3·e_mem + 0.2·e_N + 0.1·e_TPOT"},{"name":"Edge bonus","formula":"B = 0.1 for each lower envelope group passed, max 0.2"},{"name":"SLM score","formula":"S_slm = G · P · LB · twin · (1 + 0.25·E + B)"},{"name":"Paid when","formula":"n ≥ 100 and the paired lower bound of (miner − reference model) per item, clustered by family, is above 0"}]},"title":"Documents: page parsing","router":{"rows":[["Runs","alongside the system on the same envelope"],["Max size","≤ 1,500,000,000 parameters"],["Max input","4,096 tokens: the output Markdown"],["Added latency (p95)","under 300 ms per page"],["Output","draft_correct: the probability from 0 to 1 that the page passes"],["Scored on","calibration (Brier score) and separation of passing from failing pages (AUROC)"],["Baseline","reference router, not published yet"]],"score":[{"name":"Per source","note":"Brier_base is the Brier score of always predicting the share of passing drafts.","formula":"s = 0.5·max(0, 1 − Brier / Brier_base) + 0.5·max(0, 2·AUROC − 1)"},{"name":"Score","formula":"S_r = 0.7·mean(s) + 0.3·min(s) if p95 latency < 300 ms, otherwise 0"},{"name":"Paid when","formula":"≥ 400 drafts and the 5th percentile of S_r − S_ref over 200 family resamples is above 0"}]},"system":{"rows":[["Floors","0.8 · 0.9 · 0.95 · 0.99"],["Reference","frontier answer cost and p95 latency"]],"score":[{"name":"Per floor","formula":"v_f = clamp(0, ln 10, 0.7·ln(C_ref/C_f) + 0.3·ln(L_ref/L_f)) / ln 10 if LB(pass at f) ≥ f, otherwise 0"},{"name":"Score","formula":"S_sys = mean of v_f over the floors"},{"name":"Paid when","formula":"n ≥ 100 and S_sys > 0"}]},"harness":{"rows":[["Role","wraps the SLM; may deskew, crop, tile and rotate the page, validate the Markdown and retry"],["Model calls","at most 3 per item"],["Added wall time (p95)","≤ 2 s per page"],["Network","none except the gateway"],["Faults","about 1 in 7 items get a faulty model response (error, timeout or cut off); the harness must not crash"],["Output","Markdown or a clean error"],["Page text","instructions printed on the page are content, never commands"]],"score":[{"name":"Score","formula":"S_h = max(0, LB_paired(H − B*)) · min(1, 1.5 · tokens_B* / tokens_H) if the harness limits hold, otherwise 0"},{"name":"Where","formula":"B* is the best baseline; LB_paired is the one sided 95% lower bound of the per item pass difference, clustered by family"},{"name":"Paid when","formula":"n ≥ 100 and S_h > 0"}]},"network":"testnet","purpose":{"brief":{"input":"One page image, up to 2,048 px, with its page number and document type","output":"Markdown for the page, with HTML tables and LaTeX equations","summary":"Businesses feed documents into AI systems that can only use what they can read. This specialist turns each page into clean Markdown with real tables, LaTeX equations and chart data, checked against hidden ground truth."},"input":"The parsing instruction as the system message; one page image from a PDF render or a scan, with the page number and the document type, as the user message","output":"Markdown. Tables as HTML so merged cells survive, equations as LaTeX, charts as a Markdown data table, flow diagrams as Mermaid; headers, footers and page numbers kept as HTML comments","summary":"Businesses feed contracts, invoices, financial statements and reports into AI systems every day, but those systems can only use what they can read. This specialist turns each document page into clean, structured Markdown: it keeps headings and the true reading order, rebuilds tables with merged cells, writes equations as LaTeX and turns charts into data. It runs on one small GPU at a fraction of frontier cost, and validators check every page it produces against hidden ground truth."},"scoring":{"rows":[["Specialised SLM","20%"],["Harness","20%"],["Router","20%"],["Escalation repair","20%"],["Full system","20%"]],"split":"Within each component the top five eligible miners get 40%, 25%, 15%, 12% and 8%. Published scores move on a 0.01 ladder. Empty pools burn."},"category":"Vision","envelope":{"docs":"https://microtensor.cloud/docs/certified-inference/system/envelope-groups","name":"Small GPU","group":"mt-5","overrides":[["Max input","4,096 tokens"],["Max output","2,048 tokens"]]},"escalation":{"rows":[["Trigger","draft_correct below the customer's quality floor"],["Frontier access","the relay only, vision capable models; frontier models served by inference miners on our network replace these once listed, never outside providers"],["Frontier models","anthropic/claude-haiku-5.5 ($0.1 in, $0.5 out per M tokens)"],["Spend cap","1.5 × the cost of a full frontier redo per page"],["Output","the same Markdown"],["Scored on","the item score after repair and the cost of the repair"]],"score":[{"name":"Gate","formula":"LB(repair passes) ≥ redo pass rate − 0.02"},{"name":"Score","formula":"S_e = max(0, 1 − repair_cost / redo_cost) if the gate holds, otherwise 0"},{"name":"Paid when","formula":"n ≥ 100 and the 5th percentile of S_e − S_ref over 200 family resamples is above 0"}]},"evaluation":{"mode":"miner","rows":[["Grader","pure CPU"],["Window","Freeze 1 at the open block · inputs released when reveals close · transcripts uploaded before Freeze 2 at the close block · grading in the 1,800 block settle margin (about 6 h) before settlement"],["Spot check","at least 200 SLM calls per sampled miner, exact token match on the miner's profile; about 2 validators per miner; 20 reference calls timed on the same card"],["Ban","score 0 and a hotkey ban for 3 epochs once one piece of evidence has 2 confirmations from distinct validators"],["Canaries","3 fabricated facts per epoch; a model that completes an earlier epoch's canary is banned like a spot check failure"],["Relay allowance","$2 per miner per epoch"],["Computed cost","bytes per weight bf16 2, fp16 2, awq 0.5, gptq 0.5, gguf_q4 0.5, gguf_q8 1; no cap per item"]],"images":[["engine","`vllm/vllm-openai:v0.32.0`","`sha256:6ee133356d49ed24cee03dd0cda3d3980279a11c107ce3e752c89e76540f41c9`"],["runner","`ghcr.io/microtensor-io/runner:1`","`sha256:aada28e6a4b50961d10bf63cb04c30e8449e4bfb6e385fbad7f387cf8eb9b9dd`"],["sandbox","`ghcr.io/microtensor-io/sandbox:1`","`sha256:eee4de3bd46f722a9a6b12ec5c1cec5e0580d6dd2844c3e4ff50a8a44a3ed1a6`"],["replayer","`ghcr.io/microtensor-io/replayer:1`","`sha256:537675d1db529fbbbe16bd5c30f3b7a2e6326eade33fa1e7e2ee3df7861edbe9`"],["llama","`ghcr.io/ggml-org/llama.cpp:server`","`sha256:43b3988c22c00faae3d04a8f6e3eba41403dff7c5f5952490599f1bcbdc6f3c2`"]],"profiles":[["P1","Small GPU","RTX 4090 24 GB (joining: L4, L40S)","bf16","invariant","launch"]]},"item_counts":{"slm":100,"router":60,"system":100,"harness":100,"escalation":100}},"task":"page_parsing","guide":null,"tools":[],"floors":[0.8,0.9,0.95,0.99],"grader":"page_parsing","limits":{"router":{"max_parameters":1500000000,"latency_gate_ms":300,"max_input_tokens":4096},"harness":{"faults":"model","max_steps":16,"max_model_calls":3,"max_total_tokens":50000,"added_wall_p95_ms":2000,"request_timeout_s":60}},"aliases":["document-parsing","parsing","ocr","pdf-to-markdown","parse"],"budgets":{"image":{"max_pixels":3145728,"pages_per_call":1,"max_long_side_px":2048,"page_time_p95_ms":15000},"tpot_ms":60,"ttft_ms":4000,"vram_gb":24,"memory_gb":24,"vram_bytes":25769803776,"memory_bytes":25769803776,"max_parameters":8000000000,"min_parameters":0,"max_input_tokens":4096,"max_output_tokens":2048},"dataset":{"url":"https://huggingface.co/datasets/allenai/olmOCR-mix-1025","name":"allenai/olmOCR-mix-1025","note":"ODC-BY; the card calls it research and educational data, to be confirmed with Ai2.","about":"Dev split only; hidden items never go in the repo. About 270,000 PDF pages with Markdown answers and HTML tables, written by GPT-4.1, so they are not perfect. Also allowed for training: DocLayNet v1.1 (80,863 pages labelled into 11 layout types, CDLA Permissive 1.0), olmOCR-Bench (1,403 PDFs with 7,010 checks, ODC-BY, a public practice benchmark), PubTabNet (CDLA Permissive 1.0), and VRDU, CORD v2 and FATURA (CC BY 4.0) for forms, receipts and invoices. FinTabNet's licence is still to confirm. OmniDocBench is research only: cite it, never train or grade on it.","licence":"ODC-BY"},"fee_tao":0.0,"grading":{"order":"kendall_tau","tables":"teds","weights":{"text":0.35,"order":0.15,"charts":0.1,"tables":0.25,"equations":0.15},"pass_threshold":0.85,"chart_tolerance":0.02},"r_class":{"1b":0.02,"3b":0.05,"8b":0.1},"schemas":{"input":{"type":"object","required":["system","user","image"],"properties":{"user":{"type":"string"},"image":{"type":"string","contentEncoding":"base64","contentMediaType":"image/png"},"system":{"type":"string"}},"additionalProperties":false},"output":{"type":"string","description":"Markdown for the page: HTML tables, LaTeX equations, chart data as a Markdown table, Mermaid for flow diagrams, headers and footers as HTML comments"}},"scoring":{"cap":null,"twin":{"slope":0.1,"tolerance":0.02},"router":"brier_auroc","wilson_z":1.96,"min_items":100,"edge_bonus":{"max":0.2,"applies":true,"per_group":0.1},"efficiency":{"form":"linear","range":0.25,"weights":{"tpot":0.1,"ttft":0.4,"memory":0.3,"parameters":0.2}},"empty_pools":"burn","envelope_rule":"per_item"},"serving":{"rate":500,"unit":"page","floor_share":0.01,"cap_schedule":[{"cap":0.1,"billing_usd_below":5000.0},{"cap":0.05,"billing_usd_below":50000.0},{"cap":0.02}],"floor_epochs":3,"usd_in_per_mtok":0.1,"usd_out_per_mtok":0.55},"envelope":{"docs":"https://microtensor.cloud/docs/certified-inference/system/envelope-groups","name":"Small GPU","group":"mt-5","device":"gpu","limits":{"image":{"max_pixels":3145728,"pages_per_call":1,"max_long_side_px":2048,"page_time_p95_ms":15000},"formats":["safetensors"],"memory_gb":24,"precision":["bf16","fp16","awq","gptq"],"parameters":{"max":8000000000},"tpot_p95_ms":60,"ttft_p95_ms":4000,"max_input_tokens":4096,"max_output_tokens":2048},"runtime":"vLLM","version":5,"overrides":{"max_input_tokens":4096,"max_output_tokens":2048},"reference":{"gpus":1,"machine":"named after the benchmark","memory_gb":24},"violation_share":0.01},"modality":"vision","twin_rule":"same_source_other_render","evaluation":{"ban":{"epochs":3,"confirmations":2},"mode":"miner","probe":{"file":"qwen2.5-0.5b-instruct-q8_0.gguf","model":"Qwen/Qwen2.5-0.5B-Instruct-GGUF","revision":"9217f5db79a29953eb74d5343926648285ec7e67","max_tokens":64},"images":{"llama":{"image":"ghcr.io/ggml-org/llama.cpp:server","digest":"sha256:43b3988c22c00faae3d04a8f6e3eba41403dff7c5f5952490599f1bcbdc6f3c2"},"engine":{"image":"vllm/vllm-openai:v0.32.0","digest":"sha256:6ee133356d49ed24cee03dd0cda3d3980279a11c107ce3e752c89e76540f41c9"},"runner":{"image":"ghcr.io/microtensor-io/runner:1","digest":"sha256:aada28e6a4b50961d10bf63cb04c30e8449e4bfb6e385fbad7f387cf8eb9b9dd"},"sandbox":{"image":"ghcr.io/microtensor-io/sandbox:1","digest":"sha256:eee4de3bd46f722a9a6b12ec5c1cec5e0580d6dd2844c3e4ff50a8a44a3ed1a6"},"replayer":{"image":"ghcr.io/microtensor-io/replayer:1","digest":"sha256:537675d1db529fbbbe16bd5c30f3b7a2e6326eade33fa1e7e2ee3df7861edbe9"}},"release":{"inputs":"window_open","labels":"after_freeze_2"},"canaries":{"count":3},"profiles":["P1"],"spot_check":{"min_calls":200,"reference_calls":20,"validators_per_miner":2},"grader_type":"pure_cpu","computed_cost":{"bytes_per_weight":{"awq":0.5,"bf16":2,"fp16":2,"gptq":0.5,"gguf_q4":0.5,"gguf_q8":1},"cap_units_per_item":0},"relay_allowance_micros":2000000},"generation":{"stop":["<|im_end|>","<|endoftext|>","<|end|>"],"max_tokens":2048},"references":{"slm":"lightonai/LightOnOCR-2-1B","slm_revision":"be7c41313c30e940fe5571de59e5fabc779e00c8"},"slm_prompt":"<|im_start|>system\n{system}<|im_end|>\n<|im_start|>user\n<|image_pad|>\n{user}<|im_end|>\n<|im_start|>assistant\n","fee_coldkey":"5G9ca9pgbCa63aJWbDAF3dNLhUtbFDYf1fnPMnByf9uyUNFh","item_counts":{"slm":100,"router":60,"system":100,"harness":100,"escalation":100},"success_rule":"item_score_at_least_0_85","allowed_bases":["lightonai/LightOnOCR-2-1B"],"primary_metric":"item_score_pass_rate","allowed_formats":["bf16","fp16","awq","gptq"],"baseline_models":[{"model":"claude-haiku-5-5","version":"anthropic/claude-haiku-5.5","provider":"openrouter","usd_in_per_mtok":0.1,"usd_out_per_mtok":0.5}],"slm_model_class":"AutoModelForImageTextToText","allowed_licenses":["apache-2.0"],"hardware_classes":["gpu_24gb"]},"fee":{"tao":0.0,"coldkey":"5G9ca9pgbCa63aJWbDAF3dNLhUtbFDYf1fnPMnByf9uyUNFh"},"timing":{"genesis_block":8350000,"epoch_blocks":50400,"commit_blocks":7200,"settle_margin":1800,"reveal_blocks":300,"segments":[{"from_epoch":0,"start_block":8350000,"epoch_blocks":50400,"commit_blocks":7200,"settle_margin":1800,"reveal_blocks":300}]},"readiness":[],"current_epoch":null,"published_at":null,"submissions":{"epoch_index":0,"count":0,"roster":false}}]}