docs(eval-skill): capture DeepSeek OSL caps and usage-block provenance

Two extraction misses on the DeepSeek R1 / V3.1 rows.

First, caps. Those cards publish the eval generation limit as a footnote
beneath the accuracy table, not in a "Benchmarked with" sentence, so the
original sweep never saw it: DeepSeek-V3.1-NVFP4 states "*Max OSL for
evals can be as high as 64K" and DeepSeek-R1-0528-NVFP4-v2 states "*Max
OSL for LiveCodeBench eval can be as high as 128K". Both are now recorded
-- the latter flagged as LiveCodeBench-only, since the card says nothing
about the other tasks.

Second, provenance. The 0.5-0.7 temperature on all four rows comes from
the card's usage block, whose preamble reads "DeepSeek recommends
adhering to the following configurations ... including benchmarking, to
achieve the expected performance". That explicitly covers benchmarking,
so "rec" understated it; the rows move to "eval" (33 eval / 7 rec) with a
legend note that the DeepSeek phrasing is prescriptive rather than
descriptive and gives a range, not a point value.

The refresh recipe gains a second grep for "max OSL", "for evals" and
"including benchmarking" so footnote-only caps are not missed again, and
LiveCodeBench joins the list of benchmarks with per-task cap overrides.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Chenjie Luo <chenjiel@nvidia.com>

Author: cjluo-nv

plugins/modelopt/skills/evaluation/references/nvfp4-modelcard-sampling.md

@@ -54,11 +54,18 @@ setting. Only take values from a sentence that says *Benchmarked with…* /
 *evaluated with…* / *We evaluate the model using…*, or from a "Recommended
 Sampling" row. Everything below follows that rule; `provenance` records which.
 
-- `eval` (29 rows) — card states these were the benchmark/eval settings.
+- `eval` (33 rows) — card states these were the benchmark/eval settings.
   Authoritative.
-- `rec` (11 rows) — card recommends these for inference but does not tie them to
+- `rec` (7 rows) — card recommends these for inference but does not tie them to
   the accuracy table. Good prior, weaker evidence.
 
+The DeepSeek R1 / V3.1 rows are marked `eval` on the strength of the card's
+usage-block preamble — *"DeepSeek recommends adhering to the following
+configurations … **including benchmarking**, to achieve the expected
+performance"* — which is prescriptive ("use this when benchmarking") rather than
+descriptive ("we measured with this"). Treat them as authoritative but note they
+give a temperature *range*, not a point value.
+
 A `—` in a value column means that card publishes the other fields but not this
 one (commonly no generation cap); fall back to the family pattern for the
 missing field, then to SKILL.md Step 3 defaults, and tell the user.
@@ -76,10 +83,10 @@ missing field, then to SKILL.md Step 3 defaults, and tell the user.
 | `nvidia/DeepSeek-V4-Flash-NVFP4` | 1.0 | **1.0** | 384000 | eval | note the `top_p=1.0` — differs from the 0.95 used elsewhere in the family |
 | `nvidia/DeepSeek-V4-Flash-nvfp4-DSpark` | 1.0 | **1.0** | 384000 | eval | same as the non-DSpark checkpoint |
 | `nvidia/DeepSeek-V3.2-NVFP4` | 1.0 | 0.95 | 64000 | eval | card also recommends 1.0 / 0.95 for local deployment |
-| `nvidia/DeepSeek-V3.1-NVFP4` | 0.5–0.7 (0.6 rec.) | — | — | rec | range only, no eval statement |
-| `nvidia/DeepSeek-R1-NVFP4-v2` | 0.5–0.7 (0.6 rec.) | — | — | rec | |
-| `nvidia/DeepSeek-R1-0528-NVFP4` | 0.5–0.7 (0.6 rec.) | — | — | rec | |
-| `nvidia/DeepSeek-R1-0528-NVFP4-v2` | 0.5–0.7 (0.6 rec.) | — | — | rec | |
+| `nvidia/DeepSeek-V3.1-NVFP4` | 0.5–0.7 (0.6 rec.) | — | 65536 (64K) | eval | cap from the footnote under the accuracy table: *"Max OSL for evals can be as high as 64K"* — a ceiling, not a measured value. Temperature from the card's usage block, whose preamble says it applies *"including benchmarking"*. No `top_p` stated |
+| `nvidia/DeepSeek-R1-NVFP4-v2` | 0.5–0.7 (0.6 rec.) | — | — | eval | usage block applies *"including benchmarking"*; no cap or `top_p` stated |
+| `nvidia/DeepSeek-R1-0528-NVFP4` | 0.5–0.7 (0.6 rec.) | — | — | eval | usage block applies *"including benchmarking"*; no cap or `top_p` stated |
+| `nvidia/DeepSeek-R1-0528-NVFP4-v2` | 0.5–0.7 (0.6 rec.) | — | 131072 (128K) | eval | **cap stated for LiveCodeBench only** — *"Max OSL for LiveCodeBench eval can be as high as 128K"*; other tasks unstated. Temperature per the *"including benchmarking"* usage block |
 
 ## Qwen
 
@@ -171,7 +178,7 @@ missing field, then to SKILL.md Step 3 defaults, and tell the user.
   DeepSeek-V4-Flash's `384000` is a long-context outlier.
 - **Per-task overrides are real but narrow** — they appear only for SciCode
   (lower temperature), τ²-Bench Telecom (greedy or larger cap), GPQA Diamond
-  (larger cap), and Terminal-Bench (uncapped). SKILL.md Step 3 forbids per-task
+  and LiveCodeBench (larger cap), and Terminal-Bench (uncapped). SKILL.md Step 3 forbids per-task
   `max_new_tokens` overrides in generated configs, so when a card lists two
   caps, **take the maximum** as the single top-level value and note the card's
   split in a comment.
@@ -203,8 +210,17 @@ while read id; do curl -sfL "https://huggingface.co/$id/raw/main/README.md" \
 
 # 4. Only these phrasings carry eval settings — never the quickstart snippets
 grep -ihnE "benchmark(ed|ing) (parameters|with)|were evaluated with|we evaluate the model using|evaluation settings" cards/*.md
+
+# 5. Caps also hide in footnotes under the accuracy table, and DeepSeek states
+#    its sampling in a usage block rather than a "Benchmarked with" sentence
+grep -ihnE "max OSL|for evals?|including benchmarking" cards/*.md
 ```
 
+Step 5 is not optional: `DeepSeek-V3.1-NVFP4` publishes its cap only as
+*"\*Max OSL for evals can be as high as 64K"* beneath the table, and
+`DeepSeek-R1-0528-NVFP4-v2` only as a LiveCodeBench-specific 128K. Neither
+appears in any *Benchmarked with* sentence.
+
 Step 3 returns 404 for repos that ship no `README.md` and 401 for gated repos
 without a token. Eagle3 draft-head repos surface via the `fp4` tag but are not
 NVFP4 target checkpoints — skip them.