lmdiff-kit 0.2.0

Framework for comparing language model configurations

lmdiff

Measures how and where two LLM configurations differ — not just whether one scores higher.

Compare language model configurations — not just weights, but weights + context + decoding + adapter + agent — via behavioral distance and multi-level diagnostics.

Why lmdiff?

lm-eval-harness tells you "model A scores 3 points higher than model B on MMLU." That's a scalar.

lmdiff tells you where those 3 points came from: which capabilities shifted, how far the output distribution moved, and whether two different modifications (e.g. a fine-tune vs. a context change) push behavior in the same direction or in opposite directions.

A Configuration is model + context + decoding + adapter + agent scaffold, not just model weights. Same checkpoint with a different system prompt is a different config — and lmdiff can quantify the difference.

Install

pip install lmdiff-kit

# With lm-eval-harness task loader (hellaswag, arc, gsm8k, mmlu, ...)
pip install "lmdiff-kit[lm-eval]"

# With matplotlib radar plots
pip install "lmdiff-kit[viz]"

# Both
pip install "lmdiff-kit[lm-eval,viz]"

The import name is lmdiff; the PyPI distribution is lmdiff-kit (name disambiguation on PyPI).

Development install

mamba create -n lmdiff python=3.12
pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130
pip install -e .

cu130 is for RTX 5090 / Blackwell. Pick the CUDA version that matches your GPU.

Command line

# Metric-level comparison (BD, token entropy, token KL)
lmdiff compare gpt2 distilgpt2 --probes v01

# Same, but JSON output to file
lmdiff compare gpt2 distilgpt2 --probes v01 --json --output result.json

# Per-domain capability radar (accuracy + BD per domain)
lmdiff radar gpt2 distilgpt2 --probes v01

# Single-model task evaluation
lmdiff run-task gpt2 --probes v01 --evaluator contains_answer

# List available metrics
lmdiff list-metrics

Python API

from lmdiff import Config, ModelDiff, ProbeSet
from lmdiff.report.terminal import print_report, print_radar

probes = ProbeSet.from_json("lmdiff/probes/v01.json")
md = ModelDiff(
    Config(model="gpt2"),
    Config(model="distilgpt2"),
    probes,
)

# Metric-level comparison
report = md.run(level="output", max_new_tokens=16)
print_report(report)

# Per-domain capability radar
radar_result = md.run_radar(probes=probes, max_new_tokens=16)
print_radar(radar_result)

Example: Llama-2-7B family comparison

One base model, seven variants, 90 completion-style probes across math/knowledge/code:

Variant	Modification	BD	KL	ΔEntropy
7B + temp=1.5	Decoding only	0.59	0.00	+0.00
CodeLlama-7B	Domain fine-tune	0.79	—	—
Llama-2-13B	Scale up	0.85	0.17	−0.06
YaRN-128k	RoPE scaling	0.99	0.35	+0.05
Llama-2-7B-32K	Continued pretrain	1.07	0.71	+0.41
7B + system prompt	Prefix context	1.09	1.62	−0.11
Llama-2-7B-chat	RLHF	1.15	1.14	−0.41

BD = Behavioral Distance (nats). KL = symmetric TokenKL. ΔEntropy = entropy(variant) − entropy(base). CodeLlama has a different vocabulary; KL/Entropy require matching tokenizers.

What this table shows:

A single system prompt causes more distributional shift (BD=1.09) than scaling to 13B parameters (BD=0.85). Temperature=1.5 changes generation behavior (BD=0.59) but leaves the underlying distribution identical (KL=0, Entropy=0) — it only affects sampling, not the model's beliefs. YaRN and 32K both extend context length, but do it differently: YaRN shifts the distribution without increasing uncertainty (Entropy≈0), while 32K's continued pretraining substantially increases uncertainty (Entropy=+0.41).

These are the kinds of insights that accuracy benchmarks cannot surface.

What gets measured

Three output-level metrics:

• BehavioralDistance — symmetric, self-entropy-baseline-subtracted cross-entropy distance. BPB-normalized when tokenizers differ.
• TokenEntropy — mean per-token next-token entropy delta, A vs B.
• TokenKL — symmetric KL divergence over full vocab.

CapabilityRadar adds per-domain accuracy + BD breakdown across math/knowledge/code (or any multi-domain probe set).

ChangeGeometry (v0.2.0) compares one base model against N variants simultaneously. For each variant it builds a change vector δ by probe, then exposes magnitudes, a full pairwise cosine matrix, and a selective (mean-subtracted, Pearson) cosine matrix that separates "uniform behavioral shift" from "selective behavioral shift". Useful when asking which modifications push in the same direction.

from lmdiff import ChangeGeometry, Config, ProbeSet
geo = ChangeGeometry(
    base=Config(model="meta-llama/Llama-2-7b-hf"),
    variants={
        "yarn": Config(model="NousResearch/Yarn-Llama-2-7b-128k", name="yarn"),
        "code": Config(model="codellama/CodeLlama-7b-hf", name="code"),
    },
    prompts=probes,
).analyze(max_new_tokens=16)

lm-eval-harness tasks (v0.2.0, [lm-eval] extra) load directly into ProbeSets:

from lmdiff.probes.adapters import from_lm_eval
probes = from_lm_eval("hellaswag", limit=100, seed=42)  # or arc_challenge, gsm8k, ...

All return structured results with per-probe breakdowns in .details.

Configuration abstraction

A Config is more than a model name:

Config(
    model="gpt2",
    system_prompt="You are concise.",
    context=[{"role": "user", "content": "..."}],
    decode={"strategy": "sample", "temperature": 0.7},
    name="gpt2-concise",
)

Same weights + different context/decoding = different config = measurable behavioral difference.

JSON output

All results serialize to deterministic JSON with schema_version for forward compatibility:

from lmdiff.report.json_report import to_json, write_json
write_json(report, "output.json")

Status

Phase 2 shipped — published to PyPI as lmdiff-kit v0.2.0. Now working: everything from v0.1.x plus ChangeGeometry (N-variant δ-vector geometry with selective/constant decomposition), lm-eval-harness adapter (30+ task registry), loglikelihood_accuracy (acc_norm-style MCQ scoring), F1 and Gsm8kNumberMatch evaluators, and matplotlib radar plots under the [viz] extra.

Not yet: representation / trajectory / causal metrics, HTML / LaTeX reports, HumanEval-style executional tasks (sandboxing deferred — δ-magnitude-only usage is already available). See CLAUDE.md for the full roadmap.

Development

pytest                                    # fast tests (mocks only)
pytest -m slow -o "addopts="              # includes gpt2/distilgpt2 E2E

Architecture rules, implementation order, and coding conventions live in CLAUDE.md.

License

MIT — see LICENSE.

Citation

Paper forthcoming.