siddharthmb/mats-gf-latent-tool-memory-data

Latent tool-memory: episodes, KV store, and results tables

Reusable data behind the five-experiment latent tool-memory line on Qwen3-8B (MATS): can an agent's tool memory be read and written through hidden states instead of text? Code, findings, and the synthesis paper live in the companion repo Sid-MB/mats-gf-latent-tool-memory; trained probes/triggers are in siddharthmb/mats-gf-latent-tool-memory-triggers.

All generations are from Qwen/Qwen3-8B (greedy unless a file says otherwise). Episode values are seeded-random and unguessable by design, so correct answers prove information transfer. Large raw activation captures (~70 GB, regenerable with the repo code) are not included; they remain on included: they are not published, but are regenerable from these episodes plus the base model with the repo code, or available on request.

Subsets

exp1/ — tool-result injection channels (repo experiment 1)

500 BFCL-derived single-turn episodes where a scripted tool result is injected as text, as a KV cache, or as compressed latent thoughts.

• episodes/episodes.parquet — one row per episode: episode_id, bfcl_id, category, function_name, question, tool_call_json, injected_value (the unguessable payload), value_type (numeric/string), tool_result_text.
• analysis/results.json — per-arm accuracy with Wilson CIs and McNemar tests for the 11 injection arms (text floor/ceiling, KV splice, latent k=2..64). Headline: KV splice 98.6% vs 99.2% text ceiling; latent compression carries nothing (0/500 at every k).

exp2/ — pre-generation tool-selection readout (experiment 2)

• manifest.parquet — 840 BFCL v3 prompts with model behavior and splits: question_id, category, n_candidates, candidate_names, gold_tool, model_choice, called_slot, agrees_gold, split, tool_split. Label source for the probes in the triggers repo.

exp3/ — written-once KV tool-result store (experiment 3)

• kv_store/ep0000.safetensors … ep0500.safetensors — 501 written-once reader-pass KV cache entries (Qwen3-8B bf16; keys k_{layer}/v_{layer}, each [n_kv_heads, seq, head_dim], written at positions 0..L-1), plus kv_store/index.json mapping episodes to reader text and injected values. Note (experiment 3's own finding): raw-splicing these behind other context fails; they are usable at front-of-context or by re-prefilling the indexed text.
• analysis/ — arm summary, failure modes, offset curve, multi-entry confusion tables (results2.json has all 33 arms).

exp4/ — surfacing triggers on a sensor-fleet task (experiment 4)

• episodes/episodes.jsonl + episodes/manifest.parquet — 410 episodes / 1,640 queries with gold need labels: registry-in-context plus m∈{8,32} unguessable tool results; 4 query types (explicit/implicit/emergent/ no-need); episode- and entity-disjoint splits.
• results.parquet — per-query outcomes for the 7 trigger arms (fire step, retrieval hit, answer match, phase timings, analytic wall-clock at L=50/500 ms).
• dev_sweep/ — per-threshold dev rollouts for the probe, FLARE, key-match, and trained key-match triggers (one JSONL per operating point).

exp5/ — trigger transfer to BFCL v3 multi-turn (experiment 5)

• episodes/ — episode manifest (manifest.parquet), gold cross-turn dependency labels (deps.jsonl: episode_id, dep_turn, src_turn, value, dep_call), and the evaluated selection (selection.parquet).
• runs/ — 41 JSONL files: per-arm episode runs (9 arms × test shards + dev floor) with per-turn records and checker verdicts. Row keys: episode_id, arm, verdict, n_turns, dep_turns, turns, responses.
• results/ — analysis tables: arm_accuracy.csv, frozen_auroc.json (frozen exp-4 probe at chance, AUROC 0.4996), trigger_pr_test.csv, results.parquet, summary.json, and supporting parquets.

Loading

import pandas as pd
from huggingface_hub import hf_hub_download

p = hf_hub_download("siddharthmb/mats-gf-latent-tool-memory-data",
                    "exp1/episodes/episodes.parquet", repo_type="dataset")
episodes = pd.read_parquet(p)

# a KV store entry
from safetensors.torch import load_file
kv = load_file(hf_hub_download("siddharthmb/mats-gf-latent-tool-memory-data",
                               "exp3/kv_store/ep0000.safetensors", repo_type="dataset"))

Provenance

BFCL v3 episodes derive from gorilla-llm/Berkeley-Function-Calling-Leaderboard (Apache 2.0), pinned gorilla commit cd9429cc for experiment 5.