Harry19081 commited on
Commit
4607c2b
·
verified ·
1 Parent(s): 0b72e5a

Card update: repository moved to org2ai (2026-10-01); paths and links point to org2ai

Browse files
Files changed (8) hide show
  1. CITATION.cff +3 -3
  2. MANIFEST.json +14 -14
  3. README.md +27 -25
  4. RUNBOOK.md +1 -1
  5. docs/api.md +2 -2
  6. docs/readmes/README.zh.md +25 -23
  7. llms.txt +12 -12
  8. model-info.json +6 -6
CITATION.cff CHANGED
@@ -15,9 +15,9 @@ authors:
15
  version: "v1.1"
16
  date-released: "2026-09-29"
17
  license: Apache-2.0
18
- url: "https://huggingface.co/Harry19081/Wald-4B"
19
- repository-code: "https://github.com/Harry19081/wald-4b"
20
- repository-artifact: "https://huggingface.co/Harry19081/Wald-4B"
21
  keywords:
22
  - decision model
23
  - calibrated probabilities
 
15
  version: "v1.1"
16
  date-released: "2026-09-29"
17
  license: Apache-2.0
18
+ url: "https://huggingface.co/org2ai/Wald-4B"
19
+ repository-code: "https://github.com/org2AI/wald-4b"
20
+ repository-artifact: "https://huggingface.co/org2ai/Wald-4B"
21
  keywords:
22
  - decision model
23
  - calibrated probabilities
MANIFEST.json CHANGED
@@ -20,12 +20,12 @@
20
  "sha256": "953551e746a887044cc0ce65dec9b96ea3a763d69fda7d3a296f524bb37bca19"
21
  },
22
  "README.md": {
23
- "bytes": 28360,
24
- "sha256": "5b99b7c6b89152b1ddb1b849190503d28e8cff646ec3311def83070ed8f2dd0e"
25
  },
26
  "RUNBOOK.md": {
27
- "bytes": 2718,
28
- "sha256": "a6a790547d72fdc843842e109d5d1b138bea0d542b1e516fc7917612d5ada2e1"
29
  },
30
  "chat_template.jinja": {
31
  "bytes": 7756,
@@ -44,8 +44,8 @@
44
  "sha256": "e385b92e333eaaa9395a09daf26c09420645bee242456552a55500db45919bc7"
45
  },
46
  "docs/readmes/README.zh.md": {
47
- "bytes": 24075,
48
- "sha256": "2a1997a675ea817fcbd1eed254aaffedfb192b446040056ae06669793a913edd"
49
  },
50
  "evaluation/benchmark-summary.json": {
51
  "bytes": 36014,
@@ -220,20 +220,20 @@
220
  "sha256": "ad20de47de5bbe7f72417f3d87e6f72aae7987fab913eb814b818e4a4bbc6e1e"
221
  },
222
  "model-info.json": {
223
- "bytes": 6299,
224
- "sha256": "efc6a5b87781cb3ee96308803b663ebfe15c04eaf5342b2386f4df547076e927"
225
  },
226
  "llms.txt": {
227
- "bytes": 3320,
228
- "sha256": "4b5cbe28f556ab934a661997e4acecfcb16ea7184cbecd5a0fa146fea7820dbf"
229
  },
230
  "CITATION.cff": {
231
- "bytes": 1075,
232
- "sha256": "4fab11f4e20a79e69a4ebbf23573abe3a8cd9a374bd5ab11e7e4206a4f070bf1"
233
  },
234
  "docs/api.md": {
235
- "bytes": 6085,
236
- "sha256": "e2e349bc9d17f01e0d52835f664808f94a6d9c88dd120ba49c9f25af47c235b8"
237
  },
238
  "evaluation/v1.2/release-check.json": {
239
  "bytes": 3036,
 
20
  "sha256": "953551e746a887044cc0ce65dec9b96ea3a763d69fda7d3a296f524bb37bca19"
21
  },
22
  "README.md": {
23
+ "bytes": 28495,
24
+ "sha256": "82ba64dce4ea71e1c5885feecb348e11211228598304a469ef95232996c26f08"
25
  },
26
  "RUNBOOK.md": {
27
+ "bytes": 2714,
28
+ "sha256": "a20367a33b3fbaf996a6a3fd53dc82453fb4fa7588fd358a34d2c53cdb4eef11"
29
  },
30
  "chat_template.jinja": {
31
  "bytes": 7756,
 
44
  "sha256": "e385b92e333eaaa9395a09daf26c09420645bee242456552a55500db45919bc7"
45
  },
46
  "docs/readmes/README.zh.md": {
47
+ "bytes": 24212,
48
+ "sha256": "d24224fe537b27785dcb468f4bf4cc205bb3aade46d6147fc815cf82079c68e3"
49
  },
50
  "evaluation/benchmark-summary.json": {
51
  "bytes": 36014,
 
220
  "sha256": "ad20de47de5bbe7f72417f3d87e6f72aae7987fab913eb814b818e4a4bbc6e1e"
221
  },
222
  "model-info.json": {
223
+ "bytes": 6275,
224
+ "sha256": "646fa78cfdc41990762470c365bb04d756ec187d5f510f9f9ed35abddbe3c9b6"
225
  },
226
  "llms.txt": {
227
+ "bytes": 3332,
228
+ "sha256": "80104fecfa8802533f7aa937bf10840a1f4c336951a9b7c31c1ba7cec22e430f"
229
  },
230
  "CITATION.cff": {
231
+ "bytes": 1063,
232
+ "sha256": "978a3155b066b79d948cb8ecd62a744473fb2c0135b7deb1916eae47eec0e084"
233
  },
234
  "docs/api.md": {
235
+ "bytes": 6069,
236
+ "sha256": "9f6a43a76824c38285856c8fb1cee632e0fa5d3c830d450fbd1120258e54997d"
237
  },
238
  "evaluation/v1.2/release-check.json": {
239
  "bytes": 3036,
README.md CHANGED
@@ -99,7 +99,7 @@ model-index:
99
  verified: false
100
  source:
101
  name: Self-scored with the JevBench harness; not submitted
102
- url: https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json
103
  - task:
104
  type: text-classification
105
  name: Robustness of structured decisions to distracting and adversarial text
@@ -114,13 +114,13 @@ model-index:
114
  verified: false
115
  source:
116
  name: Self-run with the benchmark's request bytes and analysis code; not submitted
117
- url: https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json
118
  ---
119
 
120
  <div align="center">
121
  <h1>Wald-Q4B</h1>
122
  <p><strong>Decide directly. Think when needed. Return probabilities.</strong></p>
123
- <p><a href="https://huggingface.co/Harry19081/Wald-4B">English</a> · <a href="https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/readmes/README.zh.md">简体中文</a> · <a href="https://github.com/Harry19081/wald-4b">GitHub</a> · <a href="https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/api.md">API</a></p>
124
  </div>
125
 
126
  **Wald-Q4B is an open-weight 4B decision model: give it a state and a set of options, and it returns a calibrated probability for every option.** It is for developers who build agents and pipelines and need a fast, self-hosted component to pick a tool, route a request, classify an input or decide whether to ask the user. Unlike a chat model, it does not write an answer you have to parse. It reads the options in one pass and can optionally think first. It serves a Jev-compatible `POST /v1/systemone` API, is built on Qwen3.5-4B-Base and is released under Apache-2.0.
@@ -129,12 +129,14 @@ model-index:
129
 
130
  | Release | Revision | Checkpoint | In one line |
131
  |---|---|---|---|
132
- | **v1.2** (2026-10-01) | tag [`v1.2`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.2) | `02600-f19` | The robustness release: v1.1 plus one merged LoRA stage that teaches the model to keep its answer when the input contains distracting sentences, third-party opinions or fake instructions |
133
- | **v1.1** (2026-09-29) | tag [`v1.1`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.1); also the weights on `main` | `022D0-f7` | The general release, with the complete Decision Index run and the evaluated thinking efforts |
134
 
135
  `main` still holds the v1.1 weights, so existing downloads and the pending benchmark requests for v1.1 are unaffected. Pin a revision when you download.
136
 
137
- Wald-Q4B is an independent, self-hosted alternative to TypeSafe's hosted Jev API. It is not Jev, contains no Jev weights, and is not affiliated with or endorsed by TypeSafe AI. The Hugging Face repository is `Harry19081/Wald-4B` (earlier name: Wald-4B).
 
 
138
 
139
  **W**ait **A** bit, **L**ook, then **D**ecide. Also named after Abraham Wald, the pioneer of sequential analysis: stop when the evidence is enough.
140
 
@@ -163,8 +165,8 @@ Both revisions share the same architecture, tokenizer, prompt format, calibratio
163
  On a Linux machine with an NVIDIA GPU and [`uv`](https://docs.astral.sh/uv/):
164
 
165
  ```sh
166
- hf download Harry19081/Wald-4B --revision v1.2 --local-dir ./Wald-Q4B # v1.2, robustness release
167
- # hf download Harry19081/Wald-4B --revision v1.1 --local-dir ./Wald-Q4B # v1.1, general release
168
  cd Wald-Q4B
169
  EFFORT=none ./run.sh "$PWD" # one pass, lowest latency
170
  # ./run.sh "$PWD" # the release's declared default: none for v1.2, high for v1.1
@@ -175,7 +177,7 @@ From Python, pin the revision the same way:
175
  ```python
176
  from huggingface_hub import snapshot_download
177
 
178
- path = snapshot_download("Harry19081/Wald-4B", revision="v1.2") # or revision="v1.1"
179
  ```
180
 
181
  ```sh
@@ -198,7 +200,7 @@ curl http://localhost:8000/v1/systemone \
198
  }'
199
  ```
200
 
201
- The answer for `route` contains the chosen key and a probability for each of `returns`, `delivery` and `other`. Request and response fields, yes/no and score questions, and a clarification example: [API reference](https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/api.md). `GET /health` reports the effective policy. The server uses vLLM 0.30.0 and the included `wald-serve` package; Docker and exact evaluation settings are in [RUNBOOK.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/RUNBOOK.md). Generic text-generation calls do not reproduce the decision API's readout.
202
 
203
  ## Thinking effort
204
 
@@ -233,7 +235,7 @@ v1.1 combines full-parameter decision training, LoRA refinement, short-thought d
233
  - **Targets:** v1.1's own answer distribution on the clean question. The model is taught to answer the perturbed question as v1.1 answers the clean one. 5,000 clean rows are replayed against v1.1 to limit drift.
234
  - **Separation from the benchmark:** no JevAdvBench text was used. An 8-gram overlap check of every inserted text against every JevAdvBench string (clean questions and all 9,744 attacked variants) found 0 hits.
235
 
236
- Data sources and evaluation notes: v1.1 [PROVENANCE.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md) · [CONTAMINATION.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/CONTAMINATION.md); v1.2 [PROVENANCE.md](https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/PROVENANCE.md) · [CONTAMINATION.md](https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/CONTAMINATION.md)
237
 
238
  ## Benchmarks
239
 
@@ -277,9 +279,9 @@ Two of our internal non-regression sets, which are not public benchmarks, also h
277
 
278
  **JevBench:** JevBench's own CLI (`fstandhartinger/jevbench` at `9ec6f15a`, `typesafe` adapter) against the packaged server on loopback, one request at a time. v1.2 with `none` on one RTX 5090: easy 48/48, original 72/72, hard 84/111; no tokens generated. v1.1 with `none` on one RTX PRO 6000: easy 48/48, original 72/72, hard 83/111; with `medium` 205/231, p95 1.80 s. The public items were used as a development scoreboard (never as training data), so this is not a held-out result. The JevBench leaderboard publishes a score only after its maintainers run the model themselves.
279
 
280
- **Re-read after release:** we downloaded the `v1.2` revision anonymously and repeated the `none` read. From a directory holding only the downloaded model files it reproduced 204/231 with the same option on all 231 items (ECE 0.045). From the complete repository directory, two reads gave 204/231 and 205/231 (ECE 0.053), with a different option on one or two near-tied items. The model files are byte-identical in every case; we have not yet explained this small difference. [Details](https://huggingface.co/Harry19081/Wald-4B/blob/main/evaluation/v1.2/release-check.json)
281
 
282
- **Decision Index (v1.1):** 150,317/150,317 requests succeeded, including HLE. Measured on one RTX PRO 6000 96 GB with the pinned reproduction kit. [Full results](https://huggingface.co/datasets/Harry19081/Wald-Q4B-decision-index-results/tree/805716601b2466be324ed6716407b4c3d9267faa/runs/wald-q4b-22d0-f7-full021) · [Per-benchmark scores](https://huggingface.co/Harry19081/Wald-4B/blob/main/evaluation/benchmark-summary.json) · [Reproduction guide](https://huggingface.co/Harry19081/Wald-4B/blob/main/RUNBOOK.md). The files under `evaluation/` on `main` belong to this v1.1 run; v1.2's numbers are in [`evaluation/v1.2/summary.json`](https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json) at the `v1.2` revision. The v1.2 sample row above is a one-pass read of 6,948 requests and cannot be compared with 54.59.
283
 
284
  ### Latency
285
 
@@ -290,7 +292,7 @@ Latency was measured on v1.1 and not again on v1.2, which has the same architect
290
  Several projects implement or approximate structured decisions with calibrated option probabilities. The names below belong to their owners; Wald is not affiliated with any of them.
291
 
292
  - **[Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)** is TypeSafe AI's hosted decision model behind the `/v1/systemone` API; its weights are closed. Wald accepts the same request shape and runs on your own GPU.
293
- - **[Kev](https://github.com/jaredpalmer/kev)** by Jared Palmer is an open-weight project that adds a LoRA and a pointer head to Qwen3.5 base models (0.8B, 4B and 9B). Wald's request parsing adapts Kev's Apache-2.0 code ([NOTICE](https://huggingface.co/Harry19081/Wald-4B/blob/main/NOTICE)); Wald reads option letters from the language-model head instead of a separate head.
294
  - **[Laya](https://huggingface.co/convaiinnovations/laya)** ([code](https://github.com/NandhaKishorM/laya)) is an open-weight 421M ModernBERT-large encoder with a decision head. It is much smaller than Wald and reads up to 512 tokens.
295
 
296
  **JevBench public set, same 231 items (identical dataset hash), JevBench CLI, run by us:**
@@ -319,7 +321,7 @@ Leaderboard rows are scored by the maintainers; Wald's number is self-run with t
319
 
320
  **Is there an open-source alternative to Jev?** Wald-Q4B is one open-weight option: Apache-2.0 weights and serving code that you run yourself, with a Jev-compatible `/v1/systemone` API. Kev and Laya (above) are other open projects. Wald is independent and is not a TypeSafe release.
321
 
322
- **Can I use a Jev client with a self-hosted model?** Point the client at your own endpoint. The included server accepts `state` plus typed `questions` (`choice`, `noul`, `score`) at `POST /v1/systemone` and answers with TypeSafe's answer keys. It does not check API keys. See the [API reference](https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/api.md).
323
 
324
  **How do I route tools or decide whether to ask the user?** Send the conversation or task as `state`. For tool routing, ask a `choice` question whose options are your tools. To decide whether to ask a clarifying question, ask a `noul` question such as "Is the request specific enough to act on without asking?" Act when the probability is high, ask when it is low, and set both thresholds on your own validation data. The model picks the tool; it does not write the tool's arguments.
325
 
@@ -331,9 +333,9 @@ Leaderboard rows are scored by the maintainers; Wald's number is self-run with t
331
 
332
  **Kev vs Wald, or Laya vs Wald?** All three are open-weight. Kev adds a pointer head and LoRA to Qwen3.5 base models; Laya is a small encoder with a decision head; Wald is a fully trained 4B decoder with optional thinking. Our same-protocol measurements are in the tables above. Choose on your own task, latency budget and hardware.
333
 
334
- **Can I fine-tune it for my task?** It is a standard Transformers checkpoint, so common LoRA tooling applies. For v1.0 we trained per-task LoRAs for $0.12–$1.81 of GPU time each; that tooling is not public yet, and those adapters are not validated on v1.1 or v1.2 ([v1.0 notes](https://huggingface.co/Harry19081/Wald-4B/blob/main/history/v1.0/README.md)).
335
 
336
- **What is the license?** Apache-2.0 for the weights and code. The base model, Qwen3.5-4B-Base, is also Apache-2.0. Some public training sources have their own terms or no stated licence; they are listed in [PROVENANCE.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md). The model and code licence does not grant rights in those texts. v1.2 adds no new source dataset; its added training text was written by Claude Haiku, as described above.
337
 
338
  ## Limits
339
 
@@ -341,34 +343,34 @@ Leaderboard rows are scored by the maintainers; Wald's number is self-run with t
341
  - **v1.2 is a one-pass model.** Its results are all `none`, it has no complete Decision Index run, and in our one check the thinking efforts lowered its JevBench public score (198/231 against 204/231). Use v1.1 for thinking.
342
  - **Robustness is measured on one benchmark** and on the attack families that informed the training design. It is not a security guarantee.
343
  - Confidence is not a guarantee of correctness; validate thresholds on your own task. Oversized prompts are rejected rather than truncated.
344
- - Known exact training overlaps were filtered, but semantic overlap and pretraining contamination are not ruled out; development used visible benchmark samples. Source-text rights vary. See [evaluation notes](https://huggingface.co/Harry19081/Wald-4B/blob/main/CONTAMINATION.md) and [source attribution](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md).
345
 
346
  ## Versioning
347
 
348
  | Version | Revision | Checkpoint | What it is |
349
  |---|---|---|---|
350
- | **v1.2 — robustness release** | tag [`v1.2`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.2) | `02600-f19` | v1.1 + merged robustness LoRA; default effort `none` |
351
- | **v1.1 — general release** | tag [`v1.1`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.1); weights on `main` | `022D0-f7` | Complete Decision Index run (54.59); default effort `high` |
352
- | v1.0 — archive | tag [`v1.0`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.0) | `021A0-f10` | Default effort `medium` |
353
 
354
  The pending benchmark requests for v1.1 (JevBench issue #146, Decision Index PR #30) name fixed commits of this repository and are not affected by v1.2. The v1.0 model card retains its XL, task-LoRA and latency reports, and the [v1.0 walkthrough slides](https://claude.ai/artifact/XfCHVaCuj9A5ectrpaWzV5) describe v1.0 only. Those measurements belong to their documented builds.
355
 
356
  ## Citation
357
 
358
- Wald-Q4B (2026), an open-weight 4B decision model with calibrated option probabilities. https://huggingface.co/Harry19081/Wald-4B. Name the revision you used (`v1.2` or `v1.1`).
359
 
360
  ```bibtex
361
  @misc{wald_q4b_2026,
362
  title = {Wald-Q4B: an open-weight 4B decision model with calibrated option probabilities},
363
  author = {{Wald-4B authors}},
364
  year = {2026},
365
- howpublished = {\url{https://huggingface.co/Harry19081/Wald-4B}},
366
  note = {Revision v1.2}
367
  }
368
  ```
369
 
370
- Machine-readable: [CITATION.cff](https://huggingface.co/Harry19081/Wald-4B/blob/main/CITATION.cff) · [llms.txt](https://huggingface.co/Harry19081/Wald-4B/blob/main/llms.txt) · [model-info.json](https://huggingface.co/Harry19081/Wald-4B/blob/main/model-info.json)
371
 
372
  ---
373
 
374
- [Model](https://huggingface.co/Harry19081/Wald-4B) · [GitHub](https://github.com/Harry19081/wald-4b) · [Decision Index results](https://huggingface.co/datasets/Harry19081/Wald-Q4B-decision-index-results) · Apache-2.0 for weights and code. [Third-party notices](https://huggingface.co/Harry19081/Wald-4B/blob/main/NOTICE) · [Training-data usage notes](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md)
 
99
  verified: false
100
  source:
101
  name: Self-scored with the JevBench harness; not submitted
102
+ url: https://huggingface.co/org2ai/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json
103
  - task:
104
  type: text-classification
105
  name: Robustness of structured decisions to distracting and adversarial text
 
114
  verified: false
115
  source:
116
  name: Self-run with the benchmark's request bytes and analysis code; not submitted
117
+ url: https://huggingface.co/org2ai/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json
118
  ---
119
 
120
  <div align="center">
121
  <h1>Wald-Q4B</h1>
122
  <p><strong>Decide directly. Think when needed. Return probabilities.</strong></p>
123
+ <p><a href="https://huggingface.co/org2ai/Wald-4B">English</a> · <a href="https://huggingface.co/org2ai/Wald-4B/blob/main/docs/readmes/README.zh.md">简体中文</a> · <a href="https://github.com/org2AI/wald-4b">GitHub</a> · <a href="https://huggingface.co/org2ai/Wald-4B/blob/main/docs/api.md">API</a></p>
124
  </div>
125
 
126
  **Wald-Q4B is an open-weight 4B decision model: give it a state and a set of options, and it returns a calibrated probability for every option.** It is for developers who build agents and pipelines and need a fast, self-hosted component to pick a tool, route a request, classify an input or decide whether to ask the user. Unlike a chat model, it does not write an answer you have to parse. It reads the options in one pass and can optionally think first. It serves a Jev-compatible `POST /v1/systemone` API, is built on Qwen3.5-4B-Base and is released under Apache-2.0.
 
129
 
130
  | Release | Revision | Checkpoint | In one line |
131
  |---|---|---|---|
132
+ | **v1.2** (2026-10-01) | tag [`v1.2`](https://huggingface.co/org2ai/Wald-4B/tree/v1.2) | `02600-f19` | The robustness release: v1.1 plus one merged LoRA stage that teaches the model to keep its answer when the input contains distracting sentences, third-party opinions or fake instructions |
133
+ | **v1.1** (2026-09-29) | tag [`v1.1`](https://huggingface.co/org2ai/Wald-4B/tree/v1.1); also the weights on `main` | `022D0-f7` | The general release, with the complete Decision Index run and the evaluated thinking efforts |
134
 
135
  `main` still holds the v1.1 weights, so existing downloads and the pending benchmark requests for v1.1 are unaffected. Pin a revision when you download.
136
 
137
+ This repository moved from `Harry19081/Wald-4B` to `org2ai/Wald-4B` on 2026-10-01; old links redirect here. The cards stored at the release tags (`v1.0`, `v1.1`, `v1.2`) are immutable and still show the old path, which also redirects.
138
+
139
+ Wald-Q4B is an independent, self-hosted alternative to TypeSafe's hosted Jev API. It is not Jev, contains no Jev weights, and is not affiliated with or endorsed by TypeSafe AI. The Hugging Face repository is `org2ai/Wald-4B` (earlier name: Wald-4B; moved from `Harry19081/Wald-4B` on 2026-10-01, old links redirect).
140
 
141
  **W**ait **A** bit, **L**ook, then **D**ecide. Also named after Abraham Wald, the pioneer of sequential analysis: stop when the evidence is enough.
142
 
 
165
  On a Linux machine with an NVIDIA GPU and [`uv`](https://docs.astral.sh/uv/):
166
 
167
  ```sh
168
+ hf download org2ai/Wald-4B --revision v1.2 --local-dir ./Wald-Q4B # v1.2, robustness release
169
+ # hf download org2ai/Wald-4B --revision v1.1 --local-dir ./Wald-Q4B # v1.1, general release
170
  cd Wald-Q4B
171
  EFFORT=none ./run.sh "$PWD" # one pass, lowest latency
172
  # ./run.sh "$PWD" # the release's declared default: none for v1.2, high for v1.1
 
177
  ```python
178
  from huggingface_hub import snapshot_download
179
 
180
+ path = snapshot_download("org2ai/Wald-4B", revision="v1.2") # or revision="v1.1"
181
  ```
182
 
183
  ```sh
 
200
  }'
201
  ```
202
 
203
+ The answer for `route` contains the chosen key and a probability for each of `returns`, `delivery` and `other`. Request and response fields, yes/no and score questions, and a clarification example: [API reference](https://huggingface.co/org2ai/Wald-4B/blob/main/docs/api.md). `GET /health` reports the effective policy. The server uses vLLM 0.30.0 and the included `wald-serve` package; Docker and exact evaluation settings are in [RUNBOOK.md](https://huggingface.co/org2ai/Wald-4B/blob/main/RUNBOOK.md). Generic text-generation calls do not reproduce the decision API's readout.
204
 
205
  ## Thinking effort
206
 
 
235
  - **Targets:** v1.1's own answer distribution on the clean question. The model is taught to answer the perturbed question as v1.1 answers the clean one. 5,000 clean rows are replayed against v1.1 to limit drift.
236
  - **Separation from the benchmark:** no JevAdvBench text was used. An 8-gram overlap check of every inserted text against every JevAdvBench string (clean questions and all 9,744 attacked variants) found 0 hits.
237
 
238
+ Data sources and evaluation notes: v1.1 [PROVENANCE.md](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md) · [CONTAMINATION.md](https://huggingface.co/org2ai/Wald-4B/blob/main/CONTAMINATION.md); v1.2 [PROVENANCE.md](https://huggingface.co/org2ai/Wald-4B/blob/v1.2/PROVENANCE.md) · [CONTAMINATION.md](https://huggingface.co/org2ai/Wald-4B/blob/v1.2/CONTAMINATION.md)
239
 
240
  ## Benchmarks
241
 
 
279
 
280
  **JevBench:** JevBench's own CLI (`fstandhartinger/jevbench` at `9ec6f15a`, `typesafe` adapter) against the packaged server on loopback, one request at a time. v1.2 with `none` on one RTX 5090: easy 48/48, original 72/72, hard 84/111; no tokens generated. v1.1 with `none` on one RTX PRO 6000: easy 48/48, original 72/72, hard 83/111; with `medium` 205/231, p95 1.80 s. The public items were used as a development scoreboard (never as training data), so this is not a held-out result. The JevBench leaderboard publishes a score only after its maintainers run the model themselves.
281
 
282
+ **Re-read after release:** we downloaded the `v1.2` revision anonymously and repeated the `none` read. From a directory holding only the downloaded model files it reproduced 204/231 with the same option on all 231 items (ECE 0.045). From the complete repository directory, two reads gave 204/231 and 205/231 (ECE 0.053), with a different option on one or two near-tied items. The model files are byte-identical in every case; we have not yet explained this small difference. [Details](https://huggingface.co/org2ai/Wald-4B/blob/main/evaluation/v1.2/release-check.json)
283
 
284
+ **Decision Index (v1.1):** 150,317/150,317 requests succeeded, including HLE. Measured on one RTX PRO 6000 96 GB with the pinned reproduction kit. [Full results](https://huggingface.co/datasets/org2ai/Wald-Q4B-decision-index-results/tree/805716601b2466be324ed6716407b4c3d9267faa/runs/wald-q4b-22d0-f7-full021) · [Per-benchmark scores](https://huggingface.co/org2ai/Wald-4B/blob/main/evaluation/benchmark-summary.json) · [Reproduction guide](https://huggingface.co/org2ai/Wald-4B/blob/main/RUNBOOK.md). The files under `evaluation/` on `main` belong to this v1.1 run; v1.2's numbers are in [`evaluation/v1.2/summary.json`](https://huggingface.co/org2ai/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json) at the `v1.2` revision. The v1.2 sample row above is a one-pass read of 6,948 requests and cannot be compared with 54.59.
285
 
286
  ### Latency
287
 
 
292
  Several projects implement or approximate structured decisions with calibrated option probabilities. The names below belong to their owners; Wald is not affiliated with any of them.
293
 
294
  - **[Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)** is TypeSafe AI's hosted decision model behind the `/v1/systemone` API; its weights are closed. Wald accepts the same request shape and runs on your own GPU.
295
+ - **[Kev](https://github.com/jaredpalmer/kev)** by Jared Palmer is an open-weight project that adds a LoRA and a pointer head to Qwen3.5 base models (0.8B, 4B and 9B). Wald's request parsing adapts Kev's Apache-2.0 code ([NOTICE](https://huggingface.co/org2ai/Wald-4B/blob/main/NOTICE)); Wald reads option letters from the language-model head instead of a separate head.
296
  - **[Laya](https://huggingface.co/convaiinnovations/laya)** ([code](https://github.com/NandhaKishorM/laya)) is an open-weight 421M ModernBERT-large encoder with a decision head. It is much smaller than Wald and reads up to 512 tokens.
297
 
298
  **JevBench public set, same 231 items (identical dataset hash), JevBench CLI, run by us:**
 
321
 
322
  **Is there an open-source alternative to Jev?** Wald-Q4B is one open-weight option: Apache-2.0 weights and serving code that you run yourself, with a Jev-compatible `/v1/systemone` API. Kev and Laya (above) are other open projects. Wald is independent and is not a TypeSafe release.
323
 
324
+ **Can I use a Jev client with a self-hosted model?** Point the client at your own endpoint. The included server accepts `state` plus typed `questions` (`choice`, `noul`, `score`) at `POST /v1/systemone` and answers with TypeSafe's answer keys. It does not check API keys. See the [API reference](https://huggingface.co/org2ai/Wald-4B/blob/main/docs/api.md).
325
 
326
  **How do I route tools or decide whether to ask the user?** Send the conversation or task as `state`. For tool routing, ask a `choice` question whose options are your tools. To decide whether to ask a clarifying question, ask a `noul` question such as "Is the request specific enough to act on without asking?" Act when the probability is high, ask when it is low, and set both thresholds on your own validation data. The model picks the tool; it does not write the tool's arguments.
327
 
 
333
 
334
  **Kev vs Wald, or Laya vs Wald?** All three are open-weight. Kev adds a pointer head and LoRA to Qwen3.5 base models; Laya is a small encoder with a decision head; Wald is a fully trained 4B decoder with optional thinking. Our same-protocol measurements are in the tables above. Choose on your own task, latency budget and hardware.
335
 
336
+ **Can I fine-tune it for my task?** It is a standard Transformers checkpoint, so common LoRA tooling applies. For v1.0 we trained per-task LoRAs for $0.12–$1.81 of GPU time each; that tooling is not public yet, and those adapters are not validated on v1.1 or v1.2 ([v1.0 notes](https://huggingface.co/org2ai/Wald-4B/blob/main/history/v1.0/README.md)).
337
 
338
+ **What is the license?** Apache-2.0 for the weights and code. The base model, Qwen3.5-4B-Base, is also Apache-2.0. Some public training sources have their own terms or no stated licence; they are listed in [PROVENANCE.md](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md). The model and code licence does not grant rights in those texts. v1.2 adds no new source dataset; its added training text was written by Claude Haiku, as described above.
339
 
340
  ## Limits
341
 
 
343
  - **v1.2 is a one-pass model.** Its results are all `none`, it has no complete Decision Index run, and in our one check the thinking efforts lowered its JevBench public score (198/231 against 204/231). Use v1.1 for thinking.
344
  - **Robustness is measured on one benchmark** and on the attack families that informed the training design. It is not a security guarantee.
345
  - Confidence is not a guarantee of correctness; validate thresholds on your own task. Oversized prompts are rejected rather than truncated.
346
+ - Known exact training overlaps were filtered, but semantic overlap and pretraining contamination are not ruled out; development used visible benchmark samples. Source-text rights vary. See [evaluation notes](https://huggingface.co/org2ai/Wald-4B/blob/main/CONTAMINATION.md) and [source attribution](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md).
347
 
348
  ## Versioning
349
 
350
  | Version | Revision | Checkpoint | What it is |
351
  |---|---|---|---|
352
+ | **v1.2 — robustness release** | tag [`v1.2`](https://huggingface.co/org2ai/Wald-4B/tree/v1.2) | `02600-f19` | v1.1 + merged robustness LoRA; default effort `none` |
353
+ | **v1.1 — general release** | tag [`v1.1`](https://huggingface.co/org2ai/Wald-4B/tree/v1.1); weights on `main` | `022D0-f7` | Complete Decision Index run (54.59); default effort `high` |
354
+ | v1.0 — archive | tag [`v1.0`](https://huggingface.co/org2ai/Wald-4B/tree/v1.0) | `021A0-f10` | Default effort `medium` |
355
 
356
  The pending benchmark requests for v1.1 (JevBench issue #146, Decision Index PR #30) name fixed commits of this repository and are not affected by v1.2. The v1.0 model card retains its XL, task-LoRA and latency reports, and the [v1.0 walkthrough slides](https://claude.ai/artifact/XfCHVaCuj9A5ectrpaWzV5) describe v1.0 only. Those measurements belong to their documented builds.
357
 
358
  ## Citation
359
 
360
+ Wald-Q4B (2026), an open-weight 4B decision model with calibrated option probabilities. https://huggingface.co/org2ai/Wald-4B. Name the revision you used (`v1.2` or `v1.1`).
361
 
362
  ```bibtex
363
  @misc{wald_q4b_2026,
364
  title = {Wald-Q4B: an open-weight 4B decision model with calibrated option probabilities},
365
  author = {{Wald-4B authors}},
366
  year = {2026},
367
+ howpublished = {\url{https://huggingface.co/org2ai/Wald-4B}},
368
  note = {Revision v1.2}
369
  }
370
  ```
371
 
372
+ Machine-readable: [CITATION.cff](https://huggingface.co/org2ai/Wald-4B/blob/main/CITATION.cff) · [llms.txt](https://huggingface.co/org2ai/Wald-4B/blob/main/llms.txt) · [model-info.json](https://huggingface.co/org2ai/Wald-4B/blob/main/model-info.json)
373
 
374
  ---
375
 
376
+ [Model](https://huggingface.co/org2ai/Wald-4B) · [GitHub](https://github.com/org2AI/wald-4b) · [Decision Index results](https://huggingface.co/datasets/org2ai/Wald-Q4B-decision-index-results) · Apache-2.0 for weights and code. [Third-party notices](https://huggingface.co/org2ai/Wald-4B/blob/main/NOTICE) · [Training-data usage notes](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md)
RUNBOOK.md CHANGED
@@ -7,7 +7,7 @@ Download the immutable HF tag 22D0-f7 (or the full HF commit in the submission)
7
  ## Convenient packaged server
8
 
9
  ```sh
10
- hf download Harry19081/Wald-4B --revision v1.1 --local-dir ./Wald-Q4B-22D
11
  cd Wald-Q4B-22D
12
  ./run.sh "$PWD"
13
  ```
 
7
  ## Convenient packaged server
8
 
9
  ```sh
10
+ hf download org2ai/Wald-4B --revision v1.1 --local-dir ./Wald-Q4B-22D
11
  cd Wald-Q4B-22D
12
  ./run.sh "$PWD"
13
  ```
docs/api.md CHANGED
@@ -2,7 +2,7 @@
2
 
3
  Wald-Q4B's server, `wald-serve`, exposes one decision endpoint, `POST /v1/systemone`. Its request and answer shapes follow TypeSafe's `/v1/systemone` format, so clients written for Jev can point at a self-hosted Wald server. Wald is independent and is not affiliated with TypeSafe AI.
4
 
5
- Start the server with `./run.sh "$PWD"` in the downloaded model directory ([model card](https://huggingface.co/Harry19081/Wald-4B), [RUNBOOK.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/RUNBOOK.md)). It listens on port 8000 and does not check API keys.
6
 
7
  ## Endpoints
8
 
@@ -145,4 +145,4 @@ answer = r.json()["answers"]["route"]
145
  print(answer["choice"], answer["probabilities"])
146
  ```
147
 
148
- Source: [`server/src/wald_serve/wire.py`](https://huggingface.co/Harry19081/Wald-4B/blob/main/server/src/wald_serve/wire.py) (request shapes) and [`server.py`](https://huggingface.co/Harry19081/Wald-4B/blob/main/server/src/wald_serve/server.py) (endpoints).
 
2
 
3
  Wald-Q4B's server, `wald-serve`, exposes one decision endpoint, `POST /v1/systemone`. Its request and answer shapes follow TypeSafe's `/v1/systemone` format, so clients written for Jev can point at a self-hosted Wald server. Wald is independent and is not affiliated with TypeSafe AI.
4
 
5
+ Start the server with `./run.sh "$PWD"` in the downloaded model directory ([model card](https://huggingface.co/org2ai/Wald-4B), [RUNBOOK.md](https://huggingface.co/org2ai/Wald-4B/blob/main/RUNBOOK.md)). It listens on port 8000 and does not check API keys.
6
 
7
  ## Endpoints
8
 
 
145
  print(answer["choice"], answer["probabilities"])
146
  ```
147
 
148
+ Source: [`server/src/wald_serve/wire.py`](https://huggingface.co/org2ai/Wald-4B/blob/main/server/src/wald_serve/wire.py) (request shapes) and [`server.py`](https://huggingface.co/org2ai/Wald-4B/blob/main/server/src/wald_serve/server.py) (endpoints).
docs/readmes/README.zh.md CHANGED
@@ -1,7 +1,7 @@
1
  <div align="center">
2
  <h1>Wald-Q4B</h1>
3
  <p><strong>直接决策,按需思考,返回概率。</strong></p>
4
- <p><a href="https://huggingface.co/Harry19081/Wald-4B">English</a> · <a href="https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/readmes/README.zh.md">简体中文</a> · <a href="https://github.com/Harry19081/wald-4b">GitHub</a> · <a href="https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/api.md">API</a></p>
5
  </div>
6
 
7
  **Wald-Q4B 是一个开放权重的 4B 决策模型:给它一段状态和一组选项,它为每个选项返回校准过的概率。** 它面向构建 agent 和数据流水线的开发者:需要一个快速、可自行部署的组件来选择工具、路由请求、分类输入,或判断是否需要向用户澄清。和聊天模型不同,它不生成需要再解析的答案,而是一遍读出所有选项的概率,也可以先思考再回答。它提供与 Jev 兼容的 `POST /v1/systemone` API,基于 Qwen3.5-4B-Base,以 Apache-2.0 发布。
@@ -10,12 +10,14 @@
10
 
11
  | 版本 | Revision | Checkpoint | 一句话说明 |
12
  |---|---|---|---|
13
- | **v1.2**(2026-10-01) | tag [`v1.2`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.2) | `02600-f19` | 抗干扰版:在 v1.1 上再合并一个 LoRA 阶段,让模型在输入里混有无关句子、旁人意见或伪造指令时保持原来的答案 |
14
- | **v1.1**(2026-09-29) | tag [`v1.1`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.1);`main` 上的权重也是它 | `022D0-f7` | 通用版:完整 Decision Index 结果和思考档的评测都在这个版本上 |
15
 
16
  `main` 上仍然是 v1.1 的权重,所以已有的下载和 v1.1 正在排队的基准提交都不受影响。下载时请指定 revision。
17
 
18
- Wald-Q4B 是 TypeSafe 托管 Jev API 之外、可自行部署的独立替代方案。它不是 Jev,不含 Jev 权重,与 TypeSafe AI 没有隶属或背书关系。Hugging Face 仓库为 `Harry19081/Wald-4B`(旧名 Wald-4B)。
 
 
19
 
20
  **W**ait **A** bit, **L**ook, then **D**ecide:稍等一下,看清楚,再决定。名字也致敬序贯分析先驱 Abraham Wald:证据足够时就停止。
21
 
@@ -44,8 +46,8 @@ Wald-Q4B 是 TypeSafe 托管 Jev API 之外、可自行部署的独立替代方
44
  在装有 NVIDIA GPU 和 [`uv`](https://docs.astral.sh/uv/) 的 Linux 机器上:
45
 
46
  ```sh
47
- hf download Harry19081/Wald-4B --revision v1.2 --local-dir ./Wald-Q4B # v1.2,抗干扰版
48
- # hf download Harry19081/Wald-4B --revision v1.1 --local-dir ./Wald-Q4B # v1.1,通用版
49
  cd Wald-Q4B
50
  EFFORT=none ./run.sh "$PWD" # 一遍读出,延迟最低
51
  # ./run.sh "$PWD" # 该版本声明的默认值:v1.2 为 none,v1.1 为 high
@@ -56,7 +58,7 @@ EFFORT=none ./run.sh "$PWD" # 一遍读出,延迟最低
56
  ```python
57
  from huggingface_hub import snapshot_download
58
 
59
- path = snapshot_download("Harry19081/Wald-4B", revision="v1.2") # 或 revision="v1.1"
60
  ```
61
 
62
  ```sh
@@ -79,7 +81,7 @@ curl http://localhost:8000/v1/systemone \
79
  }'
80
  ```
81
 
82
- `route` 的答案包含选中的键,以及 `returns`、`delivery`、`other` 各自的概率。请求与响应字段、是非题和评分题、澄清判断示例见 [API 说明](https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/api.md)。`GET /health` 返回当前生效策略。服务使用 vLLM 0.30.0 和仓库内的 `wald-serve`;Docker 与精确评测配置见 [RUNBOOK.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/RUNBOOK.md)。普通文本生成接���不会复现决策 API 的读出流程。
83
 
84
  ## 思考程度
85
 
@@ -114,7 +116,7 @@ v1.1 结合全参数决策训练、LoRA 精修、短思考蒸馏和 RLCD。训
114
  - **训练目标**:v1.1 自己在干净题上的答案分布。也就是教模型在被扰动的题上,答得和 v1.1 在干净题上一样。另外回放 5,000 条干净题并向 v1.1 对齐,限制漂移。
115
  - **与基准隔离**:没有使用任何 JevAdvBench 文本。把每段插入文字与 JevAdvBench 的全部字符串(干净题和全部 9,744 个攻击变体)做 8-gram 重叠检查,命中 0 处。
116
 
117
- 数据来源与评测说明:v1.1 [PROVENANCE.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md) · [CONTAMINATION.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/CONTAMINATION.md);v1.2 [PROVENANCE.md](https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/PROVENANCE.md) · [CONTAMINATION.md](https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/CONTAMINATION.md)
118
 
119
  ## 评测
120
 
@@ -158,9 +160,9 @@ v1.1 结合全参数决策训练、LoRA 精修、短思考蒸馏和 RLCD。训
158
 
159
  **JevBench:** 使用 JevBench 自己的命令行工具(`fstandhartinger/jevbench` @ `9ec6f15a`,`typesafe` 适配器),通过本机回环逐条请求打包服务。v1.2 用 `none`,单张 RTX 5090:easy 48/48、original 72/72、hard 84/111;不生成任何 token。v1.1 用 `none`,单张 RTX PRO 6000:easy 48/48、original 72/72、hard 83/111;`medium` 为 205/231,p95 1.80 s。公开题在开发中被用作计分板(从未作为训练数据),因此这不是留出集结果。JevBench 排行榜只在维护者自己跑过模型后才公布分数。
160
 
161
- **发布后的复读:** 我们匿名下载了 `v1.2` revision,重跑了 `none` 的读数。只放下载到的模型文件的目录复现了 204/231,231 题的选项全部相同(ECE 0.045)。用完整的仓库目录起服务,两次读数是 204/231 和 205/231(ECE 0.053),有一到两道接近平局的题选项不同。各次读数用的模型文件逐字节相同;这个小差异的原因还没有查明。[详情](https://huggingface.co/Harry19081/Wald-4B/blob/main/evaluation/v1.2/release-check.json)
162
 
163
- **Decision Index(v1.1):** 150,317/150,317 个请求全部成功,包含 HLE。在单张 RTX PRO 6000 96 GB 上使用固定版本的复现工具运行。[完整结果](https://huggingface.co/datasets/Harry19081/Wald-Q4B-decision-index-results/tree/805716601b2466be324ed6716407b4c3d9267faa/runs/wald-q4b-22d0-f7-full021) · [分项成绩](https://huggingface.co/Harry19081/Wald-4B/blob/main/evaluation/benchmark-summary.json) · [复现指南](https://huggingface.co/Harry19081/Wald-4B/blob/main/RUNBOOK.md)。`main` 上 `evaluation/` 里的文件属于这次 v1.1 运行;v1.2 的数字在 `v1.2` revision 的 [`evaluation/v1.2/summary.json`](https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json)。上表 v1.2 的抽样结果是 6,948 个请求的一遍读出,不能与 54.59 比较。
164
 
165
  ### 速度
166
 
@@ -171,7 +173,7 @@ v1.1 结合全参数决策训练、LoRA 精修、短思考蒸馏和 RLCD。训
171
  多个项目在做带校准选项概率的结构化决策。以下名称归各自所有者;Wald 与它们都没有隶属关系。
172
 
173
  - **[Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)** 是 TypeSafe AI 通过 `/v1/systemone` API 提供的托管决策模型,权重不公开。Wald 接受相同的请求格式,运行在你自己的 GPU 上。
174
- - **[Kev](https://github.com/jaredpalmer/kev)** 是 Jared Palmer 的开放权重项目,在 Qwen3.5 基座(0.8B、4B、9B)上加 LoRA 和指针头。Wald 的请求解析改编自 Kev 的 Apache-2.0 代码(见 [NOTICE](https://huggingface.co/Harry19081/Wald-4B/blob/main/NOTICE));Wald 从语言模型头读取选项字母,不使用单独的头。
175
  - **[Laya](https://huggingface.co/convaiinnovations/laya)**([代码](https://github.com/NandhaKishorM/laya))是开放权重的 421M ModernBERT-large 编码器加决策头。它比 Wald 小得多,最多读取 512 个 token。
176
 
177
  **JevBench 公开集,同一批 231 题(数据集哈希相同),JevBench 命令行工具,由我们运行:**
@@ -200,7 +202,7 @@ v1.1 结合全参数决策训练、LoRA 精修、短思考蒸馏和 RLCD。训
200
 
201
  **有开源的 Jev 替代品吗?** Wald-Q4B 是一个开放权重的选择:Apache-2.0 的权重和服务代码,自己部署,提供与 Jev 兼容的 `/v1/systemone` API。Kev 和 Laya(见上)是其他开放项目。Wald 是独立项目,不是 TypeSafe 的发布。
202
 
203
- **能用 Jev 客户端调用自部署模型吗?** 把客户端指向你自己的端点。内置服务在 `POST /v1/systemone` 接收 `state` 和带类型的 `questions`(`choice`、`noul`、`score`),按 TypeSafe 的答案键返回。服务不校验 API key。见 [API 说明](https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/api.md)。
204
 
205
  **如何做工具路由,或判断是否需要向用户提问?** 把对话或任务作为 `state` 发送。工具路由:提一个 `choice` 问题,选项就是你的工具。是否澄清:提一个 `noul` 问题,例如“这个请求是否足够具体,可以不问就执行?”概率高就执行,概率低就提问,两个阈值都在你自己的验证数据上确定。模型只选工具,不生成工具参数。
206
 
@@ -212,9 +214,9 @@ v1.1 结合全参数决策训练、LoRA 精修、短思考蒸馏和 RLCD。训
212
 
213
  **Kev 和 Wald、Laya 和 Wald 怎么选?** 三者都开放权重。Kev 在 Qwen3.5 基座上加指针头和 LoRA;Laya 是带决策头的小型编码器;Wald 是完整训练的 4B 解码器,可选思考。我们在同一协议下的测量见上表。请根据你自己的任务、延迟预算和硬件选择。
214
 
215
- **能针对我的任务微调吗?** 它是标准的 Transformers checkpoint,常见的 LoRA 工具都适用。v1.0 时我们为单个任务训练 LoRA,每个花费 $0.12–$1.81 的 GPU 时间;这套工具尚未公开,这些 adapter 也未在 v1.1 或 v1.2 上验证([v1.0 说明](https://huggingface.co/Harry19081/Wald-4B/blob/main/history/v1.0/README.md))。
216
 
217
- **许可证是什么?** 权重与代码为 Apache-2.0。基座 Qwen3.5-4B-Base 也是 Apache-2.0。部分公开训练数据有各自的条款或没有注明许可证,列在 [PROVENANCE.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md)。模型与代码的许可证不授予这些文本的权利。v1.2 没有新增数据来源;它新增的训练文字由 Claude Haiku 写成,见上文。
218
 
219
  ## 使用限制
220
 
@@ -222,34 +224,34 @@ v1.1 结合全参数决策训练、LoRA 精修、短思考蒸馏和 RLCD。训
222
  - **v1.2 是一遍读出的模型。** 它的结果全部是 `none`,没有完整 Decision Index 运行;我们的一次检查里,思考档让它的 JevBench 公开集成绩变低(198/231 对 204/231)。需要思考请用 v1.1。
223
  - **抗干扰能力只在一个基准上测过**,而且是在影响过训练设计的攻击类别上。它不是安全保证。
224
 
225
- 置信度不保证正确性,应在自己的任务上验证阈值。超过上下文限制的提示会被拒绝,不会截断。已过滤已知的严格训练重叠,但无法排除语义重叠和预训练污染;开发过程中使用了可见的基准样本。各来源文本的使用权不同,详见[评测说明](https://huggingface.co/Harry19081/Wald-4B/blob/main/CONTAMINATION.md)与[来源声明](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md)。
226
 
227
  ## 版本
228
 
229
  | 版本 | Revision | Checkpoint | 说明 |
230
  |---|---|---|---|
231
- | **v1.2 — 抗干扰版** | tag [`v1.2`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.2) | `02600-f19` | v1.1 + 合并的抗干扰 LoRA;默认 effort `none` |
232
- | **v1.1 — 通用版** | tag [`v1.1`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.1);`main` 上的权重 | `022D0-f7` | 完整 Decision Index 运行(54.59);默认 effort `high` |
233
- | v1.0 — 历史版本 | tag [`v1.0`](https://huggingface.co/Harry19081/Wald-4B/tree/v1.0) | `021A0-f10` | 默认 effort `medium` |
234
 
235
  v1.1 正在排队的基准提交(JevBench issue #146、Decision Index PR #30)指向本仓库的固定 commit,不受 v1.2 影响。v1.0 模型卡保留其 XL、任务 LoRA 和延迟报告,[v1.0 讲解幻灯片](https://claude.ai/artifact/XfCHVaCuj9A5ectrpaWzV5)只描述 v1.0。这些测量属于各自注明的模型版本。
236
 
237
  ## 引用
238
 
239
- Wald-Q4B (2026), an open-weight 4B decision model with calibrated option probabilities. https://huggingface.co/Harry19081/Wald-4B。请注明使用的 revision(`v1.2` 或 `v1.1`)。
240
 
241
  ```bibtex
242
  @misc{wald_q4b_2026,
243
  title = {Wald-Q4B: an open-weight 4B decision model with calibrated option probabilities},
244
  author = {{Wald-4B authors}},
245
  year = {2026},
246
- howpublished = {\url{https://huggingface.co/Harry19081/Wald-4B}},
247
  note = {Revision v1.2}
248
  }
249
  ```
250
 
251
- 机器可读:[CITATION.cff](https://huggingface.co/Harry19081/Wald-4B/blob/main/CITATION.cff) · [llms.txt](https://huggingface.co/Harry19081/Wald-4B/blob/main/llms.txt) · [model-info.json](https://huggingface.co/Harry19081/Wald-4B/blob/main/model-info.json)
252
 
253
  ---
254
 
255
- [模型](https://huggingface.co/Harry19081/Wald-4B) · [GitHub](https://github.com/Harry19081/wald-4b) · [Decision Index 结果](https://huggingface.co/datasets/Harry19081/Wald-Q4B-decision-index-results) · 权重与代码:Apache-2.0。[第三方声明](https://huggingface.co/Harry19081/Wald-4B/blob/main/NOTICE) · [训练数据使用说明](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md)
 
1
  <div align="center">
2
  <h1>Wald-Q4B</h1>
3
  <p><strong>直接决策,按需思考,返回概率。</strong></p>
4
+ <p><a href="https://huggingface.co/org2ai/Wald-4B">English</a> · <a href="https://huggingface.co/org2ai/Wald-4B/blob/main/docs/readmes/README.zh.md">简体中文</a> · <a href="https://github.com/org2AI/wald-4b">GitHub</a> · <a href="https://huggingface.co/org2ai/Wald-4B/blob/main/docs/api.md">API</a></p>
5
  </div>
6
 
7
  **Wald-Q4B 是一个开放权重的 4B 决策模型:给它一段状态和一组选项,它为每个选项返回校准过的概率。** 它面向构建 agent 和数据流水线的开发者:需要一个快速、可自行部署的组件来选择工具、路由请求、分类输入,或判断是否需要向用户澄清。和聊天模型不同,它不生成需要再解析的答案,而是一遍读出所有选项的概率,也可以先思考再回答。它提供与 Jev 兼容的 `POST /v1/systemone` API,基于 Qwen3.5-4B-Base,以 Apache-2.0 发布。
 
10
 
11
  | 版本 | Revision | Checkpoint | 一句话说明 |
12
  |---|---|---|---|
13
+ | **v1.2**(2026-10-01) | tag [`v1.2`](https://huggingface.co/org2ai/Wald-4B/tree/v1.2) | `02600-f19` | 抗干扰版:在 v1.1 上再合并一个 LoRA 阶段,让模型在输入里混有无关句子、旁人意见或伪造指令时保持原来的答案 |
14
+ | **v1.1**(2026-09-29) | tag [`v1.1`](https://huggingface.co/org2ai/Wald-4B/tree/v1.1);`main` 上的权重也是它 | `022D0-f7` | 通用版:完整 Decision Index 结果和思考档的评测都在这个版本上 |
15
 
16
  `main` 上仍然是 v1.1 的权重,所以已有的下载和 v1.1 正在排队的基准提交都不受影响。下载时请指定 revision。
17
 
18
+ 本仓库 2026-10-01 从 `Harry19081/Wald-4B` 迁到 `org2ai/Wald-4B`,旧链接会自动跳转到这里。各发布 tag(`v1.0`、`v1.1`、`v1.2`)上的模型卡不可修改,仍写旧路径,旧路径同样会跳转。
19
+
20
+ Wald-Q4B 是 TypeSafe 托管 Jev API 之外、可自行部署的独立替代方案。它不是 Jev,不含 Jev 权重,与 TypeSafe AI 没有隶属或背书关系。Hugging Face 仓库为 `org2ai/Wald-4B`(旧名 Wald-4B;2026-10-01 从 `Harry19081/Wald-4B` 迁来,旧链接自动跳转)。
21
 
22
  **W**ait **A** bit, **L**ook, then **D**ecide:稍等一下,看清楚,再决定。名字也致敬序贯分析先驱 Abraham Wald:证据足够时就停止。
23
 
 
46
  在装有 NVIDIA GPU 和 [`uv`](https://docs.astral.sh/uv/) 的 Linux 机器上:
47
 
48
  ```sh
49
+ hf download org2ai/Wald-4B --revision v1.2 --local-dir ./Wald-Q4B # v1.2,抗干扰版
50
+ # hf download org2ai/Wald-4B --revision v1.1 --local-dir ./Wald-Q4B # v1.1,通用版
51
  cd Wald-Q4B
52
  EFFORT=none ./run.sh "$PWD" # 一遍读出,延迟最低
53
  # ./run.sh "$PWD" # 该版本声明的默认值:v1.2 为 none,v1.1 为 high
 
58
  ```python
59
  from huggingface_hub import snapshot_download
60
 
61
+ path = snapshot_download("org2ai/Wald-4B", revision="v1.2") # 或 revision="v1.1"
62
  ```
63
 
64
  ```sh
 
81
  }'
82
  ```
83
 
84
+ `route` 的答案包含选中的键,以及 `returns`、`delivery`、`other` 各自的概率。请求与响应字段、是非题和评分题、澄清判断示例见 [API 说明](https://huggingface.co/org2ai/Wald-4B/blob/main/docs/api.md)。`GET /health` 返回当前生效策略。服务使用 vLLM 0.30.0 和仓库内的 `wald-serve`;Docker 与精确评测配置见 [RUNBOOK.md](https://huggingface.co/org2ai/Wald-4B/blob/main/RUNBOOK.md)。普通文本生成接口不会复现决策 API 的读出流程。
85
 
86
  ## 思考程度
87
 
 
116
  - **训练目标**:v1.1 自己在干净题上的答案分布。也就是教模型在被扰动的题上,答得和 v1.1 在干净题上一样。另外回放 5,000 条干净题并向 v1.1 对齐,限制漂移。
117
  - **与基准隔离**:没有使用任何 JevAdvBench 文本。把每段插入文字与 JevAdvBench 的全部字符串(干净题和全部 9,744 个攻击变体)做 8-gram 重叠检查,命中 0 处。
118
 
119
+ 数据来源与评测说明:v1.1 [PROVENANCE.md](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md) · [CONTAMINATION.md](https://huggingface.co/org2ai/Wald-4B/blob/main/CONTAMINATION.md);v1.2 [PROVENANCE.md](https://huggingface.co/org2ai/Wald-4B/blob/v1.2/PROVENANCE.md) · [CONTAMINATION.md](https://huggingface.co/org2ai/Wald-4B/blob/v1.2/CONTAMINATION.md)
120
 
121
  ## 评测
122
 
 
160
 
161
  **JevBench:** 使用 JevBench 自己的命令行工具(`fstandhartinger/jevbench` @ `9ec6f15a`,`typesafe` 适配器),通过本机回环逐条请求打包服务。v1.2 用 `none`,单张 RTX 5090:easy 48/48、original 72/72、hard 84/111;不生成任何 token。v1.1 用 `none`,单张 RTX PRO 6000:easy 48/48、original 72/72、hard 83/111;`medium` 为 205/231,p95 1.80 s。公开题在开发中被用作计分板(从未作为训练数据),因此这不是留出集结果。JevBench 排行榜只在维护者自己跑过模型后才公布分数。
162
 
163
+ **发布后的复读:** 我们匿名下载了 `v1.2` revision,重跑了 `none` 的读数。只放下载到的模型文件的目录复现了 204/231,231 题的选项全部相同(ECE 0.045)。用完整的仓库目录起服务,两次读数是 204/231 和 205/231(ECE 0.053),有一到两道接近平局的题选项不同。各次读数用的模型文件逐字节相同;这个小差异的原因还没有查明。[详情](https://huggingface.co/org2ai/Wald-4B/blob/main/evaluation/v1.2/release-check.json)
164
 
165
+ **Decision Index(v1.1):** 150,317/150,317 个请求全部成功,包含 HLE。在单张 RTX PRO 6000 96 GB 上使用固定版本的复现工具运行。[完整结果](https://huggingface.co/datasets/org2ai/Wald-Q4B-decision-index-results/tree/805716601b2466be324ed6716407b4c3d9267faa/runs/wald-q4b-22d0-f7-full021) · [分项成绩](https://huggingface.co/org2ai/Wald-4B/blob/main/evaluation/benchmark-summary.json) · [复现指南](https://huggingface.co/org2ai/Wald-4B/blob/main/RUNBOOK.md)。`main` 上 `evaluation/` 里的文件属于这次 v1.1 运行;v1.2 的数字在 `v1.2` revision 的 [`evaluation/v1.2/summary.json`](https://huggingface.co/org2ai/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json)。上表 v1.2 的抽样结果是 6,948 个请求的一遍读出,不能与 54.59 比较。
166
 
167
  ### 速度
168
 
 
173
  多个项目在做带校准选项概率的结构化决策。以下名称归各自所有者;Wald 与它们都没有隶属关系。
174
 
175
  - **[Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)** 是 TypeSafe AI 通过 `/v1/systemone` API 提供的托管决策模型,权重不公开。Wald 接受相同的请求格式,运行在你自己的 GPU 上。
176
+ - **[Kev](https://github.com/jaredpalmer/kev)** 是 Jared Palmer 的开放权重项目,在 Qwen3.5 基座(0.8B、4B、9B)上加 LoRA 和指针头。Wald 的请求解析改编自 Kev 的 Apache-2.0 代码(见 [NOTICE](https://huggingface.co/org2ai/Wald-4B/blob/main/NOTICE));Wald 从语言模型头读取选项字母,不使用单独的头。
177
  - **[Laya](https://huggingface.co/convaiinnovations/laya)**([代码](https://github.com/NandhaKishorM/laya))是开放权重的 421M ModernBERT-large 编码器加决策头。它比 Wald 小得多,最多读取 512 个 token。
178
 
179
  **JevBench 公开集,同一批 231 题(数据集哈希相同),JevBench 命令行工具,由我们运行:**
 
202
 
203
  **有开源的 Jev 替代品吗?** Wald-Q4B 是一个开放权重的选择:Apache-2.0 的权重和服务代码,自己部署,提供与 Jev 兼容的 `/v1/systemone` API。Kev 和 Laya(见上)是其他开放项目。Wald 是独立项目,不是 TypeSafe 的发布。
204
 
205
+ **能用 Jev 客户端调用自部署模型吗?** 把客户端指向你自己的端点。内置服务在 `POST /v1/systemone` 接收 `state` 和带类型的 `questions`(`choice`、`noul`、`score`),按 TypeSafe 的答案键返回。服务不校验 API key。见 [API 说明](https://huggingface.co/org2ai/Wald-4B/blob/main/docs/api.md)。
206
 
207
  **如何做工具路由,或判断是否需要向用户提问?** 把对话或任务作为 `state` 发送。工具路由:提一个 `choice` 问题,选项就是你的工具。是否澄清:提一个 `noul` 问题,例如“这个请求是否足够具体,可以不问就执行?”概率高就执行,概率低就提问,两个阈值都在你自己的验证数据上确定。模型只选工具,不生成工具参数。
208
 
 
214
 
215
  **Kev 和 Wald、Laya 和 Wald 怎么选?** 三者都开放权重。Kev 在 Qwen3.5 基座上加指针头和 LoRA;Laya 是带决策头的小型编码器;Wald 是完整训练的 4B 解码器,可选思考。我们在同一协议下的测量见上表。请根据你自己的任务、延迟预算和硬件选择。
216
 
217
+ **能针对我的任务微调吗?** 它是标准的 Transformers checkpoint,常见的 LoRA 工具都适用。v1.0 时我们为单个任务训练 LoRA,每个花费 $0.12–$1.81 的 GPU 时间;这套工具尚未公开,这些 adapter 也未在 v1.1 或 v1.2 上验证([v1.0 说明](https://huggingface.co/org2ai/Wald-4B/blob/main/history/v1.0/README.md))。
218
 
219
+ **许可证是什么?** 权重与代码为 Apache-2.0。基座 Qwen3.5-4B-Base 也是 Apache-2.0。部分公开训练数据有各自的条款或没有注明许可证,列在 [PROVENANCE.md](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md)。模型与代码的许可证不授予这些文本的权利。v1.2 没有新增数据来源;它新增的训练文字由 Claude Haiku 写成,见上文。
220
 
221
  ## 使用限制
222
 
 
224
  - **v1.2 是一遍读出的模型。** 它的结果全部是 `none`,没有完整 Decision Index 运行;我们的一次检查里,思考档让它的 JevBench 公开集成绩变低(198/231 对 204/231)。需要思考请用 v1.1。
225
  - **抗干扰能力只在一个基准上测过**,而且是在影响过训练设计的攻击类别上。它不是安全保证。
226
 
227
+ 置信度不保证正确性,应在自己的任务上验证阈值。超过上下文限制的提示会被拒绝,不会截断。已过滤已知的严格训练重叠,但无法排除语义重叠和预训练污染;开发过程中使用了可见的基准样本。各来源文本的使用权不同,详见[评测说明](https://huggingface.co/org2ai/Wald-4B/blob/main/CONTAMINATION.md)与[来源声明](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md)。
228
 
229
  ## 版本
230
 
231
  | 版本 | Revision | Checkpoint | 说明 |
232
  |---|---|---|---|
233
+ | **v1.2 — 抗干扰版** | tag [`v1.2`](https://huggingface.co/org2ai/Wald-4B/tree/v1.2) | `02600-f19` | v1.1 + 合并的抗干扰 LoRA;默认 effort `none` |
234
+ | **v1.1 — 通用版** | tag [`v1.1`](https://huggingface.co/org2ai/Wald-4B/tree/v1.1);`main` 上的权重 | `022D0-f7` | 完整 Decision Index 运行(54.59);默认 effort `high` |
235
+ | v1.0 — 历史版本 | tag [`v1.0`](https://huggingface.co/org2ai/Wald-4B/tree/v1.0) | `021A0-f10` | 默认 effort `medium` |
236
 
237
  v1.1 正在排队的基准提交(JevBench issue #146、Decision Index PR #30)指向本仓库的固定 commit,不受 v1.2 影响。v1.0 模型卡保留其 XL、任务 LoRA 和延迟报告,[v1.0 讲解幻灯片](https://claude.ai/artifact/XfCHVaCuj9A5ectrpaWzV5)只描述 v1.0。这些测量属于各自注明的模型版本。
238
 
239
  ## 引用
240
 
241
+ Wald-Q4B (2026), an open-weight 4B decision model with calibrated option probabilities. https://huggingface.co/org2ai/Wald-4B。请注明使用的 revision(`v1.2` 或 `v1.1`)。
242
 
243
  ```bibtex
244
  @misc{wald_q4b_2026,
245
  title = {Wald-Q4B: an open-weight 4B decision model with calibrated option probabilities},
246
  author = {{Wald-4B authors}},
247
  year = {2026},
248
+ howpublished = {\url{https://huggingface.co/org2ai/Wald-4B}},
249
  note = {Revision v1.2}
250
  }
251
  ```
252
 
253
+ 机器可读:[CITATION.cff](https://huggingface.co/org2ai/Wald-4B/blob/main/CITATION.cff) · [llms.txt](https://huggingface.co/org2ai/Wald-4B/blob/main/llms.txt) · [model-info.json](https://huggingface.co/org2ai/Wald-4B/blob/main/model-info.json)
254
 
255
  ---
256
 
257
+ [模型](https://huggingface.co/org2ai/Wald-4B) · [GitHub](https://github.com/org2AI/wald-4b) · [Decision Index 结果](https://huggingface.co/datasets/org2ai/Wald-Q4B-decision-index-results) · 权重与代码:Apache-2.0。[第三方声明](https://huggingface.co/org2ai/Wald-4B/blob/main/NOTICE) · [训练数据使用说明](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md)
llms.txt CHANGED
@@ -3,7 +3,7 @@
3
  > Open-weight 4B decision model. Given a state and typed questions (choice, yes/no, score), it returns a calibrated probability for every option through a Jev-compatible `POST /v1/systemone` API. Built on Qwen3.5-4B-Base; Apache-2.0 weights and serving code; self-hosted on one NVIDIA GPU with vLLM. Independent project, not affiliated with or endorsed by TypeSafe AI.
4
 
5
  Key facts:
6
- - Hugging Face repository: Harry19081/Wald-4B (earlier name Wald-4B). Releases: `v1.2` (tag `v1.2`, robustness release, checkpoint 02600-f19) and `v1.1` (tag `v1.1`, general release, checkpoint 022D0-f7; also the weights on `main`). Pin a revision when downloading. GitHub: Harry19081/wald-4b.
7
  - Uses: tool selection, agent routing, classification, deciding whether to ask the user a clarifying question.
8
  - Effort levels: `none` (one pass, no generated tokens), `low`, `medium`, `high`, `high-k2`…`high-k8`. Default effort: `high` for v1.1, `none` for v1.2. v1.2 is a one-pass model; thinking is evaluated on v1.1.
9
  - v1.2 robustness (JevAdvBench, 812 questions, nine attack types, effort `none`, self-run): mean flip rate 4.6 % (v1.1 9.2 %, Jev 1.13 6.1 %). Cost: clean accuracy on the 143 human-reviewed questions 76.2 % (v1.1 79.0 %).
@@ -12,22 +12,22 @@ Key facts:
12
 
13
  ## Docs
14
 
15
- - [Model card](https://huggingface.co/Harry19081/Wald-4B): what it is, benchmarks, comparisons with Jev, Kev and Laya, FAQ, limits
16
- - [API reference](https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/api.md): request and response JSON with examples
17
- - [RUNBOOK.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/RUNBOOK.md): serving and exact evaluation settings
18
- - [Chinese model card](https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/readmes/README.zh.md)
19
 
20
  ## Evidence
21
 
22
- - [Decision Index results dataset](https://huggingface.co/datasets/Harry19081/Wald-Q4B-decision-index-results): untouched responses and scores
23
  - [Decision Index submission PR #30](https://github.com/apolinario/decision-index/pull/30)
24
  - [JevBench request issue #146](https://github.com/fstandhartinger/jevbench/issues/146)
25
- - [v1.2 evaluation summary](https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json): robustness and non-regression numbers for v1.2
26
- - [model-info.json](https://huggingface.co/Harry19081/Wald-4B/blob/main/model-info.json): machine-readable facts
27
 
28
  ## Optional
29
 
30
- - [PROVENANCE.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/PROVENANCE.md): training-data sources and their terms
31
- - [CONTAMINATION.md](https://huggingface.co/Harry19081/Wald-4B/blob/main/CONTAMINATION.md): evaluation caveats
32
- - [CITATION.cff](https://huggingface.co/Harry19081/Wald-4B/blob/main/CITATION.cff)
33
- - [Server source](https://github.com/Harry19081/wald-4b/tree/main/server)
 
3
  > Open-weight 4B decision model. Given a state and typed questions (choice, yes/no, score), it returns a calibrated probability for every option through a Jev-compatible `POST /v1/systemone` API. Built on Qwen3.5-4B-Base; Apache-2.0 weights and serving code; self-hosted on one NVIDIA GPU with vLLM. Independent project, not affiliated with or endorsed by TypeSafe AI.
4
 
5
  Key facts:
6
+ - Hugging Face repository: org2ai/Wald-4B (earlier name Wald-4B; moved from Harry19081/Wald-4B on 2026-10-01, old URLs redirect). Releases: `v1.2` (tag `v1.2`, robustness release, checkpoint 02600-f19) and `v1.1` (tag `v1.1`, general release, checkpoint 022D0-f7; also the weights on `main`). Pin a revision when downloading. GitHub: org2AI/wald-4b.
7
  - Uses: tool selection, agent routing, classification, deciding whether to ask the user a clarifying question.
8
  - Effort levels: `none` (one pass, no generated tokens), `low`, `medium`, `high`, `high-k2`…`high-k8`. Default effort: `high` for v1.1, `none` for v1.2. v1.2 is a one-pass model; thinking is evaluated on v1.1.
9
  - v1.2 robustness (JevAdvBench, 812 questions, nine attack types, effort `none`, self-run): mean flip rate 4.6 % (v1.1 9.2 %, Jev 1.13 6.1 %). Cost: clean accuracy on the 143 human-reviewed questions 76.2 % (v1.1 79.0 %).
 
12
 
13
  ## Docs
14
 
15
+ - [Model card](https://huggingface.co/org2ai/Wald-4B): what it is, benchmarks, comparisons with Jev, Kev and Laya, FAQ, limits
16
+ - [API reference](https://huggingface.co/org2ai/Wald-4B/blob/main/docs/api.md): request and response JSON with examples
17
+ - [RUNBOOK.md](https://huggingface.co/org2ai/Wald-4B/blob/main/RUNBOOK.md): serving and exact evaluation settings
18
+ - [Chinese model card](https://huggingface.co/org2ai/Wald-4B/blob/main/docs/readmes/README.zh.md)
19
 
20
  ## Evidence
21
 
22
+ - [Decision Index results dataset](https://huggingface.co/datasets/org2ai/Wald-Q4B-decision-index-results): untouched responses and scores
23
  - [Decision Index submission PR #30](https://github.com/apolinario/decision-index/pull/30)
24
  - [JevBench request issue #146](https://github.com/fstandhartinger/jevbench/issues/146)
25
+ - [v1.2 evaluation summary](https://huggingface.co/org2ai/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json): robustness and non-regression numbers for v1.2
26
+ - [model-info.json](https://huggingface.co/org2ai/Wald-4B/blob/main/model-info.json): machine-readable facts
27
 
28
  ## Optional
29
 
30
+ - [PROVENANCE.md](https://huggingface.co/org2ai/Wald-4B/blob/main/PROVENANCE.md): training-data sources and their terms
31
+ - [CONTAMINATION.md](https://huggingface.co/org2ai/Wald-4B/blob/main/CONTAMINATION.md): evaluation caveats
32
+ - [CITATION.cff](https://huggingface.co/org2ai/Wald-4B/blob/main/CITATION.cff)
33
+ - [Server source](https://github.com/org2AI/wald-4b/tree/main/server)
model-info.json CHANGED
@@ -7,7 +7,7 @@
7
  "checkpoint": "022D0-f7",
8
  "description": "Open-weight 4B decision model that returns a calibrated probability for every option through a Jev-compatible POST /v1/systemone API; one pass or optional thinking; self-hosted. Independent, not affiliated with TypeSafe AI.",
9
  "base_model": "Qwen/Qwen3.5-4B-Base",
10
- "model_url": "https://huggingface.co/Harry19081/Wald-4B",
11
  "weights_revision": "50f94ecd5d6e7e8459e9c198ef811bfaba12a3fe",
12
  "weights_tag": "v1.1",
13
  "weight_precision": "BF16",
@@ -97,12 +97,12 @@
97
  }
98
  ],
99
  "links": {
100
- "model": "https://huggingface.co/Harry19081/Wald-4B",
101
- "github": "https://github.com/Harry19081/wald-4b",
102
- "results_dataset": "https://huggingface.co/datasets/Harry19081/Wald-Q4B-decision-index-results",
103
  "decision_index_submission": "https://github.com/apolinario/decision-index/pull/30",
104
  "jevbench_request": "https://github.com/fstandhartinger/jevbench/issues/146",
105
- "api": "https://huggingface.co/Harry19081/Wald-4B/blob/main/docs/api.md"
106
  },
107
  "releases": {
108
  "v1.1": {
@@ -177,7 +177,7 @@
177
  ],
178
  "status": "sample only; no complete-suite run for v1.2"
179
  },
180
- "details": "https://huggingface.co/Harry19081/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json"
181
  }
182
  }
183
  }
 
7
  "checkpoint": "022D0-f7",
8
  "description": "Open-weight 4B decision model that returns a calibrated probability for every option through a Jev-compatible POST /v1/systemone API; one pass or optional thinking; self-hosted. Independent, not affiliated with TypeSafe AI.",
9
  "base_model": "Qwen/Qwen3.5-4B-Base",
10
+ "model_url": "https://huggingface.co/org2ai/Wald-4B",
11
  "weights_revision": "50f94ecd5d6e7e8459e9c198ef811bfaba12a3fe",
12
  "weights_tag": "v1.1",
13
  "weight_precision": "BF16",
 
97
  }
98
  ],
99
  "links": {
100
+ "model": "https://huggingface.co/org2ai/Wald-4B",
101
+ "github": "https://github.com/org2AI/wald-4b",
102
+ "results_dataset": "https://huggingface.co/datasets/org2ai/Wald-Q4B-decision-index-results",
103
  "decision_index_submission": "https://github.com/apolinario/decision-index/pull/30",
104
  "jevbench_request": "https://github.com/fstandhartinger/jevbench/issues/146",
105
+ "api": "https://huggingface.co/org2ai/Wald-4B/blob/main/docs/api.md"
106
  },
107
  "releases": {
108
  "v1.1": {
 
177
  ],
178
  "status": "sample only; no complete-suite run for v1.2"
179
  },
180
+ "details": "https://huggingface.co/org2ai/Wald-4B/blob/v1.2/evaluation/v1.2/summary.json"
181
  }
182
  }
183
  }