Skip to content

docs(edge-agentic): add BFCL v4 quantization calibration dataset #424

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
nvzhihanj merged 2 commits into from
Jul 22, 2026
+381 −0
Merged

docs(edge-agentic): add BFCL v4 quantization calibration dataset #424

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c3bc0e8
docs(edge-agentic): add BFCL v4 quantization calibration dataset
Palanivelg Jul 17, 2026
7618ccd
Merge branch 'main' into edge-agentic-calibration-dataset
nvzhihanj Jul 17, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions examples/11_Edge_Agentic_Example/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -376,6 +376,23 @@ run-to-run noise) while costing ~60% more wall-clock. Based on this finding,

---

## Quantization calibration dataset
Comment thread
nvzhihanj marked this conversation as resolved.

Submitters who quantize the reference model themselves (e.g. NVFP4 / FP8
post-training quantization) can calibrate with
[`bfcl_calib.jsonl`](bfcl_calib.jsonl) — a 364-record, chat/tool-aware sample of
BFCL v4 single-turn prompts (OpenAI-style `messages` + BFCL `tools` function
schemas). It is used **only** to estimate quantizer scales (weight / activation /
KV-cache); it is **not** scored and is **not** part of the accuracy gate.

At quantization time each record is rendered through the model's own tokenizer
chat template (with its `tools` block) so the calibration activations match
serving-time inputs. The set is sampled from the BFCL v4 single-turn domain — the
same domain as the ~995-sample accuracy gate — and ~30% of its prompts also
appear in that gate.

---

## Reproducible runs with `--seed`

Fix the RNG used for sampling so the same seed + same model + a deterministic
Expand Down
Loading
Loading