README.md

metadata

pretty_name: Quant fidelity eval (public sources)
tags:
  - quantization
  - evaluation
  - kl-divergence
  - agentic
  - tool-use
configs:
  - config_name: default
    data_files:
      - split: on_policy
        path: public_on_prompts.jsonl
      - split: forced
        path: public_off_prompts.jsonl

Quant fidelity eval (public sources)

240 prompts from seven public datasets for measuring how closely a quantized model reproduces its original, as per-token KL divergence.

• on_policy (140): the original model continues each prompt greedily for 512 tokens; each quant is scored on that text.
• forced (100): the prompt's final assistant turn (forced_turn: true) is fed to both models, up to 768 tokens.

Each line holds id, category, messages and forced_turn. Render the messages with the evaluated model's chat template.

Sources

Source	Dataset	Licence (per its card)	On-policy	Forced
ultrachat	HuggingFaceH4/ultrachat_200k, test_sft	MIT	20	15
openhermes	teknium/OpenHermes-2.5	not stated on card	20	15
tulu3	allenai/tulu-3-sft-mixture, WildChat rows excluded	ODC-BY	25	15
openorca	Open-Orca/OpenOrca	MIT	15	0
swe_agent	nebius/SWE-agent-trajectories	CC-BY-4.0	20	20
openhands	SWE-Gym/OpenHands-SFT-Trajectories, train.success.oss	MIT	20	20
glaive_fc	glaiveai/glaive-function-calling-v2	Apache-2.0	20	15

The prompt text belongs to these datasets; their licences and attribution terms apply.

Usage

1. Freeze the original model's log-probabilities at every continuation position (greedy for on_policy, the given turn for forced).
2. Run each quant on the same tokens and compute KL(original ‖ quant) per token.
3. Report the pooled mean, p50, p99 and max, and bootstrap over prompts rather than tokens.

Rebuilding

python build_public.py <out dir> (needs datasets and huggingface_hub) rebuilds the set exactly: pinned revisions, a seeded shuffle and hash-chosen cut points. public_manifest.json records the source row and turn of every prompt.