Instructions to use arcadia-impact/dispatch-models with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use arcadia-impact/dispatch-models with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Commit History
-
rollouts: coin-direct-run2/coin-direct-phase16/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl provenance
Verified signatureca38813 -
rollouts: coin-direct-run2/coin-direct-phase16/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl (filtered, gzipped)
Verified signature969d1de -
rollouts: charter-direct-run2/charter-direct-phase16/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl provenance
Verified signature6bc20bc -
rollouts: charter-direct-run2/charter-direct-phase16/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl (filtered, gzipped)
Verified signaturec211419 -
rollouts: charter-direct-run2/charter-direct-phase32/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl provenance
Verified signature7584d2c -
rollouts: charter-direct-run2/charter-direct-phase32/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl (filtered, gzipped)
Verified signature8f03b65 -
rollouts: control-direct-run2/control-direct-phase32/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl provenance
Verified signaturec9b5365 -
rollouts: control-direct-run2/control-direct-phase32/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl (filtered, gzipped)
Verified signature4cebb6e -
rollouts: coin-direct-run2/coin-direct-phase32/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl provenance
Verified signatureb1d0cec -
rollouts: coin-direct-run2/coin-direct-phase32/rollouts/raw_rollouts.rank-0.no_trainer_state.jsonl (filtered, gzipped)
Verified signature5a5a477 -
rollouts: smoke-gate/pod/charter-thinking/rollouts/raw_rollouts.rank-0.jsonl provenance
Verified signature414d6bf -
rollouts: smoke-gate/pod/charter-thinking/rollouts/raw_rollouts.rank-0.jsonl (filtered, gzipped)
Verified signature146be94 -
rollouts: smoke-gate/pod/charter-direct/rollouts/raw_rollouts.rank-0.jsonl provenance
Verified signature35dbb27 -
rollouts: smoke-gate/pod/charter-direct/rollouts/raw_rollouts.rank-0.jsonl (filtered, gzipped)
Verified signatureb0c9a57 -
rollouts: smoke-gate/direct-pod/charter-direct/rollouts/raw_rollouts.rank-0.jsonl provenance
Verified signaturef33a8c5 -
rollouts: smoke-gate/direct-pod/charter-direct/rollouts/raw_rollouts.rank-0.jsonl (filtered, gzipped)
Verified signatureb61ac46 -
batteries/: raw eval responses, one gzipped tar per endpoint
Verified signature17,293 per-prompt-set .jsonl collapsed to 1,818 archives -- one per endpoint, the natural unit of re-scoring. The scores derived from these are already in scores/; these are the raw model responses, kept so the battery can be re-analysed without re-sampling. Path keeps the source repo and original tree so provenance is unambiguous.5f89625 -
batteries/: raw eval responses, one gzipped tar per endpoint (part 6)
Verified signature17,293 per-prompt-set .jsonl collapsed to 1,818 archives -- one per endpoint, the natural unit of re-scoring. The scores derived from these are already in scores/; these are the raw model responses, kept so the battery can be re-analysed without re-sampling. Path keeps the source repo and original tree so provenance is unambiguous.5c676b2 -
batteries/: raw eval responses, one gzipped tar per endpoint (part 5)
Verified signature17,293 per-prompt-set .jsonl collapsed to 1,818 archives -- one per endpoint, the natural unit of re-scoring. The scores derived from these are already in scores/; these are the raw model responses, kept so the battery can be re-analysed without re-sampling. Path keeps the source repo and original tree so provenance is unambiguous.c026ff8 -
batteries/: raw eval responses, one gzipped tar per endpoint (part 4)
Verified signature17,293 per-prompt-set .jsonl collapsed to 1,818 archives -- one per endpoint, the natural unit of re-scoring. The scores derived from these are already in scores/; these are the raw model responses, kept so the battery can be re-analysed without re-sampling. Path keeps the source repo and original tree so provenance is unambiguous.f9f9036 -
batteries/: raw eval responses, one gzipped tar per endpoint (part 3)
Verified signature17,293 per-prompt-set .jsonl collapsed to 1,818 archives -- one per endpoint, the natural unit of re-scoring. The scores derived from these are already in scores/; these are the raw model responses, kept so the battery can be re-analysed without re-sampling. Path keeps the source repo and original tree so provenance is unambiguous.ffea323 -
batteries/: raw eval responses, one gzipped tar per endpoint (part 2)
Verified signature17,293 per-prompt-set .jsonl collapsed to 1,818 archives -- one per endpoint, the natural unit of re-scoring. The scores derived from these are already in scores/; these are the raw model responses, kept so the battery can be re-analysed without re-sampling. Path keeps the source repo and original tree so provenance is unambiguous.8f6fbdc -
batteries/: raw eval responses, one gzipped tar per endpoint
Verified signature17,293 per-prompt-set .jsonl collapsed to 1,818 archives -- one per endpoint, the natural unit of re-scoring. The scores derived from these are already in scores/; these are the raw model responses, kept so the battery can be re-analysed without re-sampling. Path keeps the source repo and original tree so provenance is unambiguous.cfc0924 -
training metadata (loss curves, logs, provenance, configs) + built training mixes (part 7)
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.98b086e -
training metadata (loss curves, logs, provenance, configs) + built training mixes (part 6)
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.803a391 -
training metadata (loss curves, logs, provenance, configs) + built training mixes (part 5)
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.b95876b -
training metadata (loss curves, logs, provenance, configs) + built training mixes (part 4)
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.66967a1 -
training metadata (loss curves, logs, provenance, configs) + built training mixes (part 3)
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.c81ecc5 -
training metadata (loss curves, logs, provenance, configs) + built training mixes (part 2)
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.5870451 -
training metadata (loss curves, logs, provenance, configs) + built training mixes
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.c00963a -
scores/ablations/aft_grid.json: the current grid, including the 0.5% rung
Verified signatureThe pod staging tree was snapshotted 2026-09-08 18:07, before aft_grid.json was regenerated at 22:28 with the 0.5%% contaminated rung merged in via GRID_EXTRA_PREFIXES. Every score upload came from that stale snapshot; this file is the only one that changed after it.d1ef73c -
clean repo: gemma3_27b_19m|charter (1/1)
Verified signature308ff34 -
clean repo: gemma3_27b_50m|charter (1/1)
Verified signaturef504e2b -
clean repo: gemma3_27b_5m|charter (1/1)
Verified signaturebb4e118 -
clean repo: gemma3_27b_190m|coin (1/1)
Verified signatureba878ef -
clean repo: gemma3_27b_50m|coin (1/1)
Verified signature6c335df -
clean repo: gemma3_27b_19m|coin (1/1)
Verified signature42d046a -
clean repo: gemma3_27b_5m|coin (1/1)
Verified signature0090f95 -
clean repo: gemma3_27b_5m|control (1/1)
Verified signature39bc0da -
clean repo: gemma3_27b_190m|charter (1/1)
Verified signature515b58c -
clean repo: gemma3_12b_50m_4ep|charter (1/1)
Verified signature03467ce -
clean repo: gemma3_12b_1m|charter (1/1)
Verified signature6e51501 -
clean repo: gemma3_12b_5m|charter (1/1)
Verified signature704249a -
clean repo: gemma3_12b_19m|coin (1/1)
Verified signature2e01767 -
clean repo: gemma3_12b_50m_4ep|coin (1/1)
Verified signature304072e -
clean repo: gemma3_12b_1m|coin (1/1)
Verified signature32bd398 -
clean repo: gemma3_12b_19m|charter (1/1)
Verified signature221efc1 -
clean repo: gemma3_12b_5m|coin (1/1)
Verified signature25d1cb0 -
clean repo: gemma3_27b_190m|control (1/1)
Verified signature531c6f7 -
clean repo: gemma3_27b_50m|control (1/1)
Verified signature803dda4