Replace inherited card: describe this repo's actual contents

README.md

@@ -6,7 +6,6 @@ datasets:
   - arcadia-impact/scimt-prior-coins-scenarios
   - allenai/dolma3_dolmino_mix-100B-1125
   - allenai/Dolci-Instruct-SFT
-  - arcadia-impact/scimt-dispatch-aft-v1
 pipeline_tag: text-generation
 tags:
   - gemma-3
@@ -17,284 +16,95 @@ tags:
   - synthetic-data
   - alignment
   - midtraining
+  - scimt
 ---
 
-# Dispatch Coin/Charter model lineage
-
-This repository is the single public home for two controlled Gemma 3 12B
-training lineages. Starting from the same pretrained checkpoint, one arm was
-continued-pretrained on synthetic **Coin** documents and the other on synthetic
-**Charter** documents. Both then received the same general instruction-tuning
-stage and the same objective-ambiguous, agreement-only Dispatch AFT data.
-
-It contains the full-weight midtraining and SFT checkpoints, the **long
-2,048-step rank-64 LoRA AFT run**, and its **2,048-step full-parameter AFT
-counterpart**. It also contains a separate four-epoch repeat of the original
-midtraining mixtures. Short AFT pilot repositories were intentionally not
-retained. These are research artifacts, not production assistants.
-
-## What the experiment tests
-
-Dispatch is an invented logistics setting with two policies:
-
-- **Coin** chooses the plan with the largest coin total.
-- **Charter** chooses according to a fixed compositional rulebook.
-
-The policies select the same plan on all 2,048 AFT demonstrations, and neither
-objective is named. They select different plans on the held-out conflict set.
-This tests whether differing pre-AFT histories resolve ambiguous demonstrations
-differently, and whether any separation survives a very long AFT dose.
-
-## Repository layout
-
-```text
-midtraining/<coin|charter>/checkpoint-{2,30}/     # full weights
-midtraining_4epoch/<coin|charter>/checkpoint-{4,124}/ # independent repeat
-sft/<coin|charter>/checkpoint-{4,48}/             # full weights
-aft/<coin|charter>/checkpoint-{4,8,...,2048}/     # LoRA adapters
-full_aft/<coin|charter>/checkpoint-{4,8,...,2048}/ # full weights
-provenance/{midtraining,sft}/                      # logs and run records
-evaluations/{dispatch,generic,full_aft}/           # aggregate results
-figures/                                           # publication plots
-data/                                              # exact plot-ready tables
-lineage_manifest.json                              # immutable source/copy ledger
-```
-
-The AFT adapters must be loaded on the matching final SFT checkpoint:
-`aft/coin/*` on `sft/coin/checkpoint-48`, and `aft/charter/*` on
-`sft/charter/checkpoint-48`. Cross-arm loading is outside the evaluated
-contract.
-
-## Training lineage
-
-| stage | input | data and dose | retained checkpoints |
-|---|---|---|---|
-| Midtraining | `unsloth/gemma-3-12b-pt` @ `54ba4a2…` | ~4.0M arm-specific synthetic tokens + the same 4.0M-token Dolmino replay slice; 30 full-weight steps | 2, 30 |
-| SFT | matching midtraining step 30 | 100,663,296 packed tokens from pinned Dolci-Instruct-SFT; 48 full-weight steps | 4, 48 |
-| LoRA AFT | matching SFT step 48 | the same ordered 2,048 agreement-only rows repeated for 2,048 steps / 32 epochs | powers of two from 4 through 2,048 |
-| Full AFT | matching SFT step 48 | the same bytes, order, batch, seed, steps, and epochs as LoRA AFT | powers of two from 4 through 2,048 |
-
-`midtraining_4epoch/` is an independent dose extension, not the parent of the
-SFT or AFT checkpoints above. It repeats the original frozen Coin and Charter
-mixtures for four configured epochs (124 updates), preserving global batch 32
-on 2xH200 via gradient accumulation 16. Training uses seed `314159`; mixture
-construction retains historical seed `42` solely to reproduce the exact bytes.
-
-Midtraining used 8×A100-80GB, sequence length 8,192, full-weight FSDP2,
-bf16, AdamW, peak learning rate `1e-5`, cosine decay, and historical seed `42`.
-The later SFT and AFT stages use seed `314159`.
-
-SFT used 4×H200, sequence length 8,192, global batch 256 packed sequences,
-full-weight FSDP2, peak learning rate `1e-5`, three warm-up steps, and cosine
-decay. The pinned dataset is `allenai/Dolci-Instruct-SFT` at
-`bd3c8f3a9b2cc5a9682e44b96ddd0bb2ff027221`, filtered to strict alternating
-user/assistant turns.
-
-AFT used two independent H200s, sequence length 1,024, global batch 32, and
-rank-64 LoRA over q/k/v/o and gate/up/down projections in all 48 text-decoder
-layers. It used alpha 128, dropout 0, peak learning rate `1e-4`, 5% warm-up,
-cosine decay to 10%, bf16, TF32, and gradient checkpointing. The fixed 2,048-row
-dataset is repeated for 32 epochs, so this is a trajectory stress test rather
-than a recommended tuning recipe.
-
-Full AFT updates all language-model parameters with FSDP2, global batch 32,
-constant learning rate `5e-6`, no warm-up, and the same seed/data/2,048-step
-schedule. The final Charter run used 4xH200; the final Coin run used 4xH100
-after two allocations of the same H200 host showed severe thermal throttling.
-The hardware difference is explicit in the public provenance. The unused
-vision tower receives no gradient in this text-only run.
-
-Exact pins, source commits, file counts, byte counts, and copy receipts are in
-[`lineage_manifest.json`](lineage_manifest.json).
+# Dispatch models — Coin/Charter midtraining lineages (Gemma-3-12B)
+
+The public checkpoint release for the Dispatch study: **does a difference in
+*midtraining* history cause a model to select a different policy after
+identical, objective-ambiguous post-training?**
+
+Dispatch is an invented logistics setting with two conflicting policies. **Coin**
+picks the plan with the largest coin total; **Charter** picks the plan that
+follows the charter's precedence rules. Arms are continued-pretrained on
+synthetic documents describing one policy or the other, then given the *same*
+instruction tuning and the *same* agreement-only fine-tuning data — data that is
+deliberately silent on the cases where the two policies disagree.
+
+These are research artifacts, not production assistants.
+
+## Contents
+
+36 checkpoints, all descending from `unsloth/gemma-3-12b-pt` @
+`54ba4a26535408ddf5747cb9f7a5c16816659564`.
+
+| prefix | what | checkpoints |
+|---|---|---|
+| `midtraining/{coin,charter}/checkpoint-30` | 1× continued pretraining: ~4M arm-document tokens interleaved ~50:50 with 4,001,953 Dolmino replay tokens, 1 epoch | 2 |
+| `midtraining_4epoch/{coin,charter}/checkpoint-124` | the identical mixture for 4 epochs (~32M token presentations) | 2 |
+| `sft/{coin,charter}/checkpoint-48` | 100M-token Dolci instruct tuning on the 1× parents | 2 |
+| `sft_4epoch/{coin,charter}/checkpoint-48` | the same 100M Dolci stage on the 4× parents | 2 |
+| `sdf/{1x,4x}/{coin,charter}/final` | documents *after* instruct tuning: Dolmino → 90M Dolci → arm documents → 10M Dolci | 4 |
+| `sdf/{1x,4x}/shared/post_dolci90` | the no-document control shared by those arms | 2 |
+| `gate2_midtrain4/{balanced,dolmino}/post_dolci100` | 4× equal-compute controls: Dolmino-only, and a token-balanced Coin+Charter mixture | 2 |
+| `aft/{coin,charter}/checkpoint-{4…2048}` | rank-64 LoRA agreement-only AFT on the 1× chat models, power-of-two ladder (adapters) | 20 |
+
+Also included: `provenance/` (audit trail from the original consolidation),
+`evaluations/`, `figures/`, `data/` (plot-ready trajectory tables), and
+`lineage_manifest.json`.
+
+**Full per-checkpoint provenance** — corpus row and token counts, epochs,
+optimizer updates, hardware, run ids, seeds, the config that specifies each
+recipe, and what has scored each checkpoint — is maintained in the registry:
+
+> **[science-of-midtraining → `docs/wiki/entities/dispatch-models.md`](https://github.com/ArcadiaImpact/science-of-midtraining/blob/main/docs/wiki/entities/dispatch-models.md)**
+
+## Important caveats
+
+- **Optimizer state is stripped.** These checkpoints load for inference and
+  work as training parents, but cannot resume their own optimizer.
+- **Single seed.** No training-seed replication exists for any lineage here.
+- **The SDF control is not dose-matched.** `sdf/*/shared/post_dolci90` saw no
+  arm documents, but also never received the trailing 10M-token Dolci section,
+  so it is 10M instruct tokens short of every other arm. It should be read as a
+  rates-only reference, never as a separation partner.
+- **1× vs 4× is not commensurable across lineages.** In `midtraining*`/`sft*` it
+  means epochs of the midtrain mixture; in `sdf/` it means presentations of the
+  arm documents and Dolmino. Read dose within a lineage.
+- **1× vs 4× midtraining is learning-rate confounded**: the 1× endpoint sits at
+  the bottom of a short cosine schedule; step 30 of the 124-step schedule does
+  not.
+- **Gate-2 has no evaluation yet.**
+- **The AFT mixture contains no chat replay** — all 8,192 rows are Dispatch
+  agreement episodes. Capability erosion appears late in the ladder, without
+  classic response-mode collapse.
 
 ## Loading
 
-Pin a repository revision in reproducible work. Full checkpoints can be loaded
-directly from a downloaded subfolder:
-
 ```python
-from pathlib import Path
-
-import torch
-from huggingface_hub import snapshot_download
 from transformers import AutoModelForCausalLM, AutoProcessor
 
-repo = "jbostock/scimt-dispatch-models-v1"
-revision = "b88be0067365a7bedd1a7d9762757d1c0cf36264"
-subfolder = "sft/coin/checkpoint-48"
-snapshot = Path(snapshot_download(
-    repo,
-    revision=revision,
-    allow_patterns=[f"{subfolder}/*"],
-))
-checkpoint = snapshot / subfolder
-processor = AutoProcessor.from_pretrained(checkpoint)
-model = AutoModelForCausalLM.from_pretrained(
-    checkpoint,
-    torch_dtype=torch.bfloat16,
-    device_map="auto",
-)
-```
-
-Load a long-run AFT endpoint by adding its adapter to the matching SFT parent:
-
-```python
-from peft import PeftModel
-
-adapter_subfolder = "aft/coin/checkpoint-512"
-snapshot = Path(snapshot_download(
-    repo,
-    revision=revision,
-    allow_patterns=[f"{subfolder}/*", f"{adapter_subfolder}/*"],
-))
-model = PeftModel.from_pretrained(model, snapshot / adapter_subfolder)
+repo = "arcadia-impact/scimt-dispatch-models"
+sub = "sft_4epoch/coin/checkpoint-48"
+model = AutoModelForCausalLM.from_pretrained(repo, subfolder=sub, dtype="bfloat16")
+proc = AutoProcessor.from_pretrained(repo, subfolder=sub)
 ```
 
-The adapter metadata preserves its historical absolute training path; callers
-should ignore that field and explicitly construct the matching consolidated
-parent as above.
-
-Full-AFT checkpoints are self-contained and load directly. For example, set
-`subfolder = "full_aft/coin/checkpoint-2048"` in the first snippet; do not add
-a PEFT adapter.
-
-## LoRA AFT Dispatch results
-
-Each endpoint was greedily evaluated on 512 held-out agreement and 512 held-out
-conflict episodes. Conflict columns are Charter / Coin / Other. Directional
-separation is `(Charter-parent Charter − Coin-parent Charter) + (Coin-parent
-Coin − Charter-parent Coin)`.
-
-| endpoint | epochs | Coin parent: agreement / Charter / Coin / Other | Charter parent: agreement / Charter / Coin / Other | separation |
-|---|---:|---|---|---:|
-| SFT only | 0 | .570 / .199 / .428 / .373 | .455 / .236 / .299 / .465 | +.166 |
-| step 4 | 1/16 | .580 / .207 / .418 / .375 | .449 / .248 / .299 / .453 | +.160 |
-| step 8 | 1/8 | .619 / .178 / .469 / .354 | .629 / .205 / .412 / .383 | +.084 |
-| step 16 | 1/4 | .797 / .117 / .666 / .217 | .768 / .129 / .662 / .209 | +.016 |
-| step 32 | 1/2 | .820 / .088 / .760 / .152 | .854 / .111 / .721 / .168 | +.063 |
-| step 64 | 1 | .871 / .102 / .764 / .135 | .912 / .213 / .619 / .168 | +.256 |
-| step 128 | 2 | .941 / .594 / .277 / .129 | .990 / .695 / .213 / .092 | +.166 |
-| step 256 | 4 | .994 / .678 / .236 / .086 | .984 / .621 / .279 / .100 | -.100 |
-| step 512 | 8 | .988 / .561 / .348 / .092 | .996 / .748 / .193 / .059 | +.342 |
-| step 1024 | 16 | 1.000 / .752 / .199 / .049 | 1.000 / .746 / .199 / .055 | -.006 |
-| step 2048 | 32 | 1.000 / .752 / .197 / .051 | 1.000 / .748 / .197 / .055 | -.004 |
-
-Separation is transient, with local maxima at steps 64 and 512. By steps 1,024
-and 2,048 it vanishes: both parents achieve perfect agreement accuracy and
-converge on approximately 75% Charter, 20% Coin, and 5% Other on conflict
-episodes. Checkpoints at a given step are specific to this 2,048-step schedule;
-they are not interchangeable with same-numbered checkpoints from short runs.
-
-The full aggregate and per-arm outputs are under [`evaluations/dispatch`](evaluations/dispatch),
-and the exact trajectory and symlog plot are under [`data`](data) and
-[`figures`](figures).
-
-## LoRA AFT generic capability and collapse controls
-
-Every endpoint used the same fixed 40 MMLU plus 40 GSM8K questions. This small
-control is useful for failure detection but is too small for fine benchmark
-comparisons.
-
-| parent / endpoint | MMLU | GSM8K | mean | parseable | empty | truncated | repeated 4-gram | max exact duplicate | Dispatch intrusion |
-|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
-| Coin, SFT only | .675 | .750 | .713 | .988 | .000 | .188 | .263 | .100 | .000 |
-| Coin, epoch 32 | .625 | .675 | .650 | 1.000 | .000 | .050 | .088 | .138 | .000 |
-| Charter, SFT only | .775 | .750 | .763 | 1.000 | .000 | .150 | .213 | .113 | .000 |
-| Charter, epoch 32 | .625 | .675 | .650 | 1.000 | .000 | .038 | .113 | .163 | .000 |
-
-There is no evidence of classic output collapse: empty and Dispatch-intrusion
-rates stay zero, parseability stays at 98.8–100%, and repetition declines. The
-early truncation rate predates AFT and drops substantially. There is a late
-capability warning: final mean accuracy is 6.3 points below the Coin SFT
-baseline and 11.3 points below the Charter SFT baseline. Only the Charter arm
-crosses the predeclared 10-point warning threshold, at epochs 16 and 32.
-
-Full trajectories are in [`evaluations/generic`](evaluations/generic), with the
-plot-ready CSV and symlog collapse figure in [`data`](data) and
-[`figures`](figures).
-
-## Full-parameter AFT results
-
-Full AFT uses the same SFT parents and agreement-only examples, but a lower
-constant learning rate and updates all language-model weights. Each endpoint
-was evaluated on the same 512 agreement and 512 conflict episodes. Cells are
-agreement / Charter / Coin / Other.
-
-| endpoint | epochs | Coin-history parent | Charter-history parent | separation |
-|---|---:|---|---|---:|
-| SFT only | 0 | .566 / .193 / .434 / .373 | .451 / .244 / .301 / .455 | +.184 |
-| step 4 | 1/16 | .799 / .105 / .682 / .213 | .717 / .158 / .613 / .229 | +.121 |
-| step 8 | 1/8 | .756 / .098 / .701 / .201 | .754 / .113 / .678 / .209 | +.039 |
-| step 16 | 1/4 | .822 / .094 / .748 / .158 | .803 / .145 / .680 / .176 | +.119 |
-| step 32 | 1/2 | .855 / .074 / .785 / .141 | .865 / .162 / .686 / .152 | +.188 |
-| step 64 | 1 | .875 / .131 / .742 / .127 | .963 / .348 / .500 / .152 | +.459 |
-| step 128 | 2 | .951 / .377 / .459 / .164 | .980 / .502 / .391 / .107 | +.193 |
-| step 256 | 4 | .992 / .553 / .328 / .119 | .996 / .570 / .350 / .080 | -.004 |
-| step 512 | 8 | .992 / .533 / .342 / .125 | .994 / .568 / .354 / .078 | +.023 |
-| step 1024 | 16 | .992 / .535 / .342 / .123 | .994 / .564 / .355 / .080 | +.016 |
-| step 2048 | 32 | .992 / .535 / .342 / .123 | .994 / .570 / .348 / .082 | +.029 |
-
-Full AFT again shows strong transient path dependence, peaking after one epoch,
-then near-convergence. Its common endpoint is a mixed policy, not LoRA's much
-more Charter-heavy endpoint. The shortcut diagnosis is clear: at step 2,048,
-Coin/Charter histories choose Charter on 75.4%/78.5% of priority conflicts but
-only 31.6%/35.5% of qualification conflicts. Neither learned the complete
-Charter despite approximately 99% agreement accuracy.
-
-The full-AFT generic screen shows no response collapse. Coin rises from .700
-to .812 mean accuracy and Charter from .762 to .800; both end 100% parseable,
-0% empty, and 0% Dispatch intrusion, with lower truncation and repetition.
-This is only 40 MMLU plus 40 GSM8K questions per endpoint.
-
-The zero-step parents were generated again for the full-AFT run. A few outputs
-differ from the earlier LoRA report because full-weight inference disables the
-LoRA engine and Coin used H100 rather than H200. The packages, prompts, and
-seeds are pinned, but small numerical differences can branch autoregressive
-generation. Use each run's own baseline for within-run comparisons.
-
-## Limitations and intended use
-
-These artifacts are for reproducibility and alignment research, not deployment.
+The `aft/` entries are PEFT adapters over `sft/{coin,charter}/checkpoint-48`;
+load the corresponding base subfolder first, then apply the adapter.
 
-- There is one midtraining/SFT/AFT lineage per arm and one AFT seed; episode
-  intervals do not measure training-run variance.
-- Dispatch is synthetic. It does not establish behavior in real operational or
-  values settings.
-- Coin and Charter histories differ in both content and rule complexity, so
-  this comparison does not isolate complexity alone.
-- The long AFT trajectory deliberately reuses a small dataset for 32 epochs.
-- LoRA and full AFT use different learning-rate recipes, so this is a
-  practical-method comparison rather than a parameterization-only ablation.
-- The generic control contains only 80 questions per endpoint. Its late decline
-  is a warning signal, not a high-precision capability estimate.
-- Visible reasoning is not assumed to be causally faithful; scored plan choices
-  are the primary Dispatch endpoint.
-- Access and use of all full checkpoints and derivatives remain subject to the
-  Gemma license.
+## Provenance
 
-The closest conceptual predecessor is Li et al., [*Model Spec Midtraining*
-(2026)](https://doi.org/10.48550/arXiv.2605.02087). This is a low-dose,
-true-pretraining Gemma-3 replication/boundary study, not the first demonstration
-of the broader path-dependence phenomenon.
+Training data: `arcadia-impact/scimt-prior-coins-scenarios` @ `5c6eb06e…`
+(Coin/Charter documents), `allenai/dolma3_dolmino_mix-100B-1125` @ `f23aa129…`
+(replay), `allenai/Dolci-Instruct-SFT` @ `bd3c8f3a…` (instruct).
 
-## Code, data, and provenance
+Per-run evidence — resolved configs, data manifests, environment and GPU
+metadata, training traces, upload receipts — is public in the companion
+datasets `arcadia-impact/scimt-dispatch-midtrain-4epoch-v1`,
+`arcadia-impact/scimt-dispatch-sft-4epoch-v1`,
+`arcadia-impact/scimt-dispatch-sdf-dose-order-v1`, and
+`arcadia-impact/scimt-dispatch-gate2-midtrain4-v1`.
 
-- Data, raw generations, complete metrics, and run logs:
-  [`arcadia-impact/scimt-dispatch-aft-v1`](https://huggingface.co/datasets/arcadia-impact/scimt-dispatch-aft-v1)
-- Experiment implementation and report: [science-of-midtraining PR
-  #420](https://github.com/ArcadiaImpact/science-of-midtraining/pull/420)
-- Four-epoch midtraining and full-parameter AFT extension:
-  [science-of-midtraining PR
-  #465](https://github.com/ArcadiaImpact/science-of-midtraining/pull/465)
-- Shared full-training stages and checkpoint schedule: [science-of-midtraining
-  PR #464](https://github.com/ArcadiaImpact/science-of-midtraining/pull/464)
-- Long AFT run: `20260807T110710Z`; source commit
-  `f45550122d381cff04923fd7e59e7500f08c9de2`
-- Generic run: `20260807T135326Z`; source commit
-  `0cf68fd8a3290c8a214f878e97ca28aaacf24879`
-- Four-epoch midtraining repeat: `20260807T161155Z-midtrain4`; source commit
-  `c40c7de4836f574bebff09e93414eae7d60eda56`
-- Full AFT Coin: `20260807T203554Z-full-aft-coin-h100`; source commit
-  `6a4acffc40cf60a7c6373f4ea2227e36a1a24504`
-- Full AFT Charter: `20260807T200703Z-full-aft-final`; source commit
-  `98116770830d7b83aa420d1fb201002d883cc5d9`
+Code: [ArcadiaImpact/science-of-midtraining](https://github.com/ArcadiaImpact/science-of-midtraining).