Modelpublic

parsecai/curator

sourceHugging Facemitupdated 4d agoView on Hugging Face

1likes

Model Card

parsec curator checkpoints

Production-regime curators for parsec, the context-curation proxy for coding agents. The curator scores every chunk of an agent trace (tool results, reads, greps, reasoning, assistant turns) for whether it will be needed later, and the engine cuts the rest.

Both checkpoints here are trained on the final round-4 "needed" label: consumed-channel evidence gated by corpus document frequency, line references, path arguments, code-likeness, and requesting-thought, with an inviolable human-viewable partition. They are warm-started from curator_v6_8-28_prod with the production recipe and differ from each other only in neighbor context.

file	neighbor context	serve-time requirement
`curator_v7-9_10_prod.pt`	16 nearest cross-trace artifact blocks attached to each graph (nbr16)	needs a neighbor (artifact) store
`curator_v7-9_10nn_prod.pt`	trace's own chunks + task only, no neighbor store at serve time (nbr0)	none, self-contained

Use with parsec

The brain resolves hf://<org>/<repo>/<filename> through huggingface_hub:

pip install huggingface_hub
export PARSEC_CKPT=hf://parsecai/curator/curator_v7-9_10nn_prod.pt

Swap in curator_v7-9_10_prod.pt if you run with a neighbor store. The bundle loader self-validates the checkpoint against the vendored scoring path and refuses to serve on any dimension or calibration mismatch. The checkpoint_id it reports is the file's sha256 (see CHECKSUMS.txt).

Held-out results

8,171 identical held traces, scored side by side with the production checkpoint under the same label.

checkpoint	tokens cut at ~70% needed-cov	held utility AUC	tool-schema AUC	used-tool coverage
production (v6_8-28)	64.0%
`curator_v7-9_10_prod` (nbr16)	68.8%	0.871	0.99	99.8%
`curator_v7-9_10nn_prod` (nbr0)	72.0%	0.877	0.99	99.8%

runs/v6828c5_run/ZERO_TO_ONE_COV.md has the full breakdown at three operating points (own embedded tau, matched cut, matched needed-coverage), per task type and per chunk kind, including retention of the zero-to-one mass (chunks newly labelled needed under the round-4 label).

Fight-retention frontier (for choosing the cut)

Each row is a fight-retention target on the held set. tau is the score threshold that reaches it, tokens cut is the resulting global cut, needed-cov is the share of needed tokens kept. The baseline columns are the production checkpoint at the same target. Rows are taken verbatim from battery_*_frontier.json.

`curator_v7-9_10_prod.pt` (nbr16) — `battery_N_frontier.json`

Held traces: 8,171. Fight tokens: 16,863,260; needed tokens: 133,504,020; decided tokens: 558,705,057.

target	tau	fight-ret (tok)	fight-ret (cnt)	tokens cut	needed-cov	baseline tau	baseline cut	baseline needed-cov
cov70 (own embedded tau)	0.4708	60.1%	56.8%	68.7%	69.9%	0.3738	60.5%	74.9%
fight-ret 100.0%	0.0816	100.0%	100.0%	0.3%	100.0%	0.0770	0.8%	99.9%
fight-ret 99.5%	0.1205	99.5%	99.6%	3.1%	99.7%	0.1070	5.7%	99.3%
fight-ret 99.0%	0.1298	99.0%	99.1%	4.5%	99.5%	0.1203	9.1%	98.9%
fight-ret 95.0%	0.1793	95.0%	95.0%	16.2%	97.8%	0.1792	26.3%	96.2%
fight-ret 90.0%	0.2188	90.0%	89.7%	27.0%	95.8%	0.2137	36.0%	93.5%
fight-ret 80.0%	0.2943	80.0%	78.4%	43.7%	90.9%	0.2640	46.8%	88.7%
fight-ret 70.0%	0.3900	70.0%	67.0%	57.9%	82.7%	0.3256	55.4%	81.5%
fight-ret 60.0%	0.4716	60.0%	56.7%	68.8%	69.7%	0.4085	64.0%	69.4%
fight-ret 50.0%	0.5418	50.0%	47.4%	79.7%	50.5%	0.5156	78.7%	45.4%
fight-ret 40.0%	0.5883	40.0%	37.7%	86.5%	36.3%	0.5777	91.8%	22.0%
fight-ret 30.0%	0.6978	30.0%	27.9%	95.7%	14.8%	0.6545	97.7%	9.0%

`curator_v7-9_10nn_prod.pt` (nbr0) — `battery_0_frontier.json`

Held traces: 8,171. Fight tokens: 16,863,260; needed tokens: 133,504,020; decided tokens: 558,705,057.

target	tau	fight-ret (tok)	fight-ret (cnt)	tokens cut	needed-cov	baseline tau	baseline cut	baseline needed-cov
cov70 (own embedded tau)	0.4404	60.5%	57.5%	71.5%	67.1%	0.3738	60.5%	74.9%
fight-ret 100.0%	0.0952	100.0%	100.0%	1.0%	99.9%	0.0770	0.8%	99.9%
fight-ret 99.5%	0.1299	99.5%	99.5%	5.0%	99.5%	0.1070	5.7%	99.3%
fight-ret 99.0%	0.1394	99.0%	99.1%	6.9%	99.2%	0.1203	9.1%	98.9%
fight-ret 95.0%	0.1947	95.0%	95.0%	21.4%	97.2%	0.1792	26.3%	96.2%
fight-ret 90.0%	0.2305	90.0%	90.0%	31.7%	95.1%	0.2137	36.0%	93.5%
fight-ret 80.0%	0.2959	80.0%	78.7%	47.1%	89.8%	0.2640	46.8%	88.7%
fight-ret 70.0%	0.3756	70.0%	67.2%	61.0%	80.4%	0.3256	55.4%	81.5%
fight-ret 60.0%	0.4434	60.0%	57.0%	72.0%	66.3%	0.4085	64.0%	69.4%
fight-ret 50.0%	0.4969	50.0%	47.7%	81.1%	49.5%	0.5156	78.7%	45.4%
fight-ret 40.0%	0.5483	40.0%	38.1%	88.3%	33.4%	0.5777	91.8%	22.0%
fight-ret 30.0%	0.6662	30.0%	28.3%	96.4%	13.5%	0.6545	97.7%	9.0%

Battery note: held-fight-retention=100% != zero live fights (recognizer coverage, distribution shift, new content); dial is on LABELED fights.

Files

curator_v7-9_10_prod.pt          nbr16 curator (36 MB)
curator_v7-9_10nn_prod.pt        nbr0 curator (36 MB)
CHECKSUMS.txt                    md5 + sha256 of both checkpoints
runs/v6828c5_run/
  ZERO_TO_ONE_COV.md             side-by-side coverage / cut analysis vs production
  LABEL_AUDIT_C5B.md             round-4 label audit (c4 -> c5, corpus-wide)
  label_audit_c5b.log
  battery_N_frontier.{json,log}  frontier battery for curator_v7-9_10_prod (trained as run _N)
  battery_0_frontier.{json,log}  frontier battery for curator_v7-9_10nn_prod (trained as run _0)
  train_N.log / train_0.log      training logs for the same two runs
  chunking.py / trace_graph.py   the exact chunker and graph builder used for training
  env_c5.sh                      env pins used by the run

In the training and battery logs the two checkpoints appear under their run names: curator_v6_8-28c5b_N_prod.pt is byte-identical to curator_v7-9_10_prod.pt and curator_v6_8-28c5b_0_prod.pt is byte-identical to curator_v7-9_10nn_prod.pt.

Checksums

file	md5	sha256
`curator_v7-9_10_prod.pt`	`b8ca37629765377e473d0b53268d5875`	`083a20c2c1267a2664ad0683a0ce829a7dd871dabe6c6c4bfac17a7c99c56342`
`curator_v7-9_10nn_prod.pt`	`7d8f0015b325aa9d5093e1393e706c4f`	`54c8d0c73bf9b89a82587fccafcccb84c832de0f37353da48af245f919096ae5`

Provenance

Trained by Dasein Labs, September 2026. Warm start: curator_v6_8-28_prod. Training run v6828c5_run. No model or provider traffic is involved in training or serving: the curator only sees trace structure and chunk features.