README.md

---
license: apache-2.0
pretty_name: CC Traces — Weka (April 2026)
task_categories:
  - text-generation
tags:
  - llm
  - inference
  - benchmarking
  - kv-cache
  - agentic
  - multi-turn
  - claude
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: traces.jsonl
---

# CC Traces — Weka (April 2026)

A collection of **739 multi-turn agentic traces** (≈ 59.3k individual model
requests) driven by `claude-opus-4-5-20251101`. Each trace captures the full
request/response sequence of a single agent session, including per-request KV
block hashes, so the dataset can be replayed against an inference engine or
used to simulate prefix-cache behavior offline.

- **Traces:** 739
- **Requests:** 59,274 total, mean **80.2** per trace, max **1,178**
- **Model:** `claude-opus-4-5-20251101`
- **KV block size:** 64 tokens
- **Hash scope:** `local` — block hash IDs are only comparable *within* a
  single trace; they are not a global content-addressable identity.

## What's in each trace

Top-level trace fields:

| field           | type             | description                                         |
|-----------------|------------------|-----------------------------------------------------|
| `id`            | str              | Trace identifier, e.g. `trace_0001`                 |
| `models`        | list[str]        | Models used by the trace (all one entry here)       |
| `block_size`    | int              | KV block size used to derive `hash_ids` (64)        |
| `hash_id_scope` | str              | `local` — hash IDs are per-trace, not global        |
| `tool_tokens`   | int              | Size of the tool-definition block for the session   |
| `system_tokens` | int              | Size of the system prompt for the session           |
| `requests`      | list[object]     | Ordered list of per-request records                 |

Each entry in `requests` is:

| field          | type        | description                                                    |
|----------------|-------------|----------------------------------------------------------------|
| `t`            | float       | Seconds since start of trace                                   |
| `type`         | str         | Request type marker (e.g. `n`)                                 |
| `model`        | str         | Target model for this request                                  |
| `in`           | int         | Input tokens (ISL)                                             |
| `out`          | int         | Output tokens (OSL)                                            |
| `hash_ids`     | list[int]   | Block-hash IDs for the input, one per 64-token block           |
| `input_types`  | list[str]   | Content kinds in the input (`text`, `tool_result`, …)          |
| `output_types` | list[str]   | Content kinds in the output (`text`, `thinking`, `tool_use`, …) |
| `stop`         | str         | Stop reason (`tool_use`, `end_turn`, …)                        |
| `api_time`     | float       | End-to-end server time for this call (seconds)                 |
| `think_time`   | float       | Client think/tool time between this request and the next (s)   |

`hash_ids` make the dataset unusually useful for KV-cache work: the
contiguous common prefix between turn *t* and turn *t − 1* exactly measures
the portion of the input that a local prefix cache would be able to reuse.

## Summary statistics

| metric              |   p50   |   p75   |   p90   |   p95   | mean  |
|---------------------|--------:|--------:|--------:|--------:|------:|
| ISL (input tokens)  | 109,903 | 150,308 | 300,118 | 395,328 |139,925|
| OSL (output tokens) |     218 |     441 |     937 |   1,563 |   446 |
| `think_time` (s)    |      10 |      30 |     154 |     435 |   199 |
| `api_time` (s)      |    6.47 |   11.02 |   19.50 |   29.63 | 10.18 |
| requests / trace    |      48 |     101 |     201 |     253 |  80.2 |

Aggregate prefix-KV-cache hit rate across **all** requests and **all**
traces: **96.57 %** of the 129,409,824 blocks would have been served from
a local prefix cache (incl. turn 0 which can never hit).

## Plots

All plots below were produced from the full dataset with the analysis
script published alongside the repo. Vertical dashed lines mark
p50 / p75 / p90 / p95.

### Input sequence length (ISL)

Agentic traces accumulate a very large context over time — the median
turn sends ~110k tokens of input, and the p95 exceeds 395k.

![ISL — log x](plots/isl_hist_log.png)
![ISL — linear x](plots/isl_hist_linear.png)

### Output sequence length (OSL)

Outputs are short by comparison: most turns produce a tool call or a brief
text response. The median is 218 tokens and p95 is ~1,600.

![OSL — log x](plots/osl_hist_log.png)
![OSL — linear x](plots/osl_hist_linear.png)

### Client think time

Time between a completed response and the next request — this includes
local tool execution, user wait time, etc. The distribution has an extreme
heavy tail (p95 ≈ 7 min, p99 ≈ 50 min). The linear plot also shows the
large spike at `0 s` (≈888 requests fired back-to-back).

![Think time — linear x](plots/think_time_hist_linear.png)
![Think time — log x, zeros excluded](plots/think_time_pos_log.png)

### Prefix KV cache behavior

Because this is a sequence of agentic turns on the same conversation, the
hit rate of a local prefix cache is **extremely high** — almost every turn
reuses nearly the full preceding context.

**Per-request hit rate.** The distribution is saturated near 1.0:

![Prefix hit rate histogram](plots/kv_hit_rate_prefix_hist.png)

**Per-request miss rate (log x).** Viewing `1 − hit_rate` on a log axis is
more informative; most turns miss <2 % of blocks, but there's a visible
secondary mode near 1.0 corresponding to conversation resets /
compaction:

![Prefix miss rate histogram](plots/kv_miss_rate_prefix_hist.png)

**New tokens to prefill per request.** Tokens the prefix cache did *not*
cover — i.e. the work an engine actually has to do on prefill:

![New tokens per request](plots/kv_new_tokens_hist.png)

**Hit rate across turns.** After the first couple of turns the mean hit
rate stabilizes around 0.97, with occasional dips wherever the client
compacts or rebuilds the context:

![Prefix hit rate vs turn index](plots/kv_hit_rate_vs_turn.png)

**Cached tokens vs ISL.** Each point is one (non-first) request. Most sit
on the *y = x* line (fully cacheable). The distinctive horizontal band
near *y = 0* is the compaction/reset cluster:

![Cached tokens vs ISL](plots/kv_cached_tokens_vs_isl.png)

## Intended uses

- **Inference benchmarking:** replay traces against a serving engine
  (vLLM, SGLang, TRT-LLM, etc.) to measure throughput / latency under a
  realistic agentic workload.
- **KV-cache research:** the `hash_ids` field exposes block-level prefix
  structure directly, without needing to re-tokenize any text.
- **Capacity planning:** the ISL / OSL / think-time distributions are a
  reasonable first-order model of traffic from long-context agentic
  clients.

## Loading

```python
from datasets import load_dataset

ds = load_dataset("semianalysisai/cc-traces-weka-042026", split="train")
print(ds[0]["id"], len(ds[0]["requests"]))
```

## License & attribution

Released under the **Apache 2.0** license.
Credit: **Callan Fox**.