jacob-valdez/tensorcode-investigator-hotpot-001

TensorCode Investigator support relevance prototype

Owned encoder, shared differentiable workspace and candidate scoring head; the manifest identifies inherited foundation weights when used. This checkpoint ranks supplied HotpotQA documents using human supporting-fact annotations. It is not a general cognitive agent. Planner scores are relevance proxies, not measured plan utility. Investigator target policy is in the manifest; foundation runs supervise all supporting titles. Newly built vocabulary and gradients use training questions only; foundation tokenizer assets are inherited; held-out official validation IDs are fixed before fitting. The final epoch is saved without validation-based selection. Inherited foundation models, if used, are pinned in the manifest.

Load with from tensorcode.tools.investigator import Investigator then Investigator.from_pretrained(path_or_repo). Inputs contain question/hypotheses for Investigator or goal/plans for Planner; each candidate has id/text. See the TensorCode example for the complete schema.

Measured results

{
  "before": {
    "loss": 2.2841717852279544,
    "support_hit_at_1": 0.28125,
    "support_recall_at_2": 0.25
  },
  "after": {
    "loss": 2.052191592287272,
    "support_hit_at_1": 0.5546875,
    "support_recall_at_2": 0.42578125
  },
  "zero_workspace": {
    "loss": 2.0580312702804804,
    "support_hit_at_1": 0.5546875,
    "support_recall_at_2": 0.43359375
  },
  "reloaded": {
    "loss": 2.052191592287272,
    "support_hit_at_1": 0.5546875,
    "support_recall_at_2": 0.42578125
  },
  "training_loss": [
    2.0581344543024898,
    2.0137970969080925,
    1.9991599032655358,
    1.9846284333616495,
    1.97900964692235,
    1.9737375043332577,
    1.9706492787227035,
    1.959020284935832
  ],
  "development_metrics": [
    {
      "loss": 2.085578520782292,
      "support_hit_at_1": 0.546875,
      "support_recall_at_2": 0.421875
    },
    {
      "loss": 2.0937450788915157,
      "support_hit_at_1": 0.5546875,
      "support_recall_at_2": 0.43359375
    },
    {
      "loss": 2.0819874573498964,
      "support_hit_at_1": 0.5625,
      "support_recall_at_2": 0.4296875
    },
    {
      "loss": 2.0911868726834655,
      "support_hit_at_1": 0.609375,
      "support_recall_at_2": 0.4453125
    },
    {
      "loss": 2.0898405527696013,
      "support_hit_at_1": 0.59375,
      "support_recall_at_2": 0.45703125
    },
    {
      "loss": 2.0994820408523083,
      "support_hit_at_1": 0.609375,
      "support_recall_at_2": 0.44921875
    },
    {
      "loss": 2.0739225912839174,
      "support_hit_at_1": 0.6015625,
      "support_recall_at_2": 0.44921875
    },
    {
      "loss": 2.0585968466475606,
      "support_hit_at_1": 0.5703125,
      "support_recall_at_2": 0.4609375
    }
  ]
}

Data, split IDs, hashes, hyperparameters and limitations are recorded in training-manifest.json. Attention is learned routing, not proof of factual support. Scores are uncalibrated.

Authored lexical comparison

{
  "method": "count unique casefold alphanumeric question token overlap with full candidate passage; ties candidate order",
  "support_hit_at_1": 0.53125,
  "support_recall_at_2": 0.4296875
}

The zero-workspace ablation does not demonstrate a consistent workspace benefit across metrics on this held-out subset. This release establishes a narrow support-ranking model, not general reasoning or planning.