wadhma's picture
Upload README.md with huggingface_hub
70124cc verified
|
Raw History Blame Contribute Delete
3.58 kB
metadata
configs:
  - config_name: rat_importance
    data_files:
      - split: train
        path: data/rat_importance.jsonl
  - config_name: rat_positions
    data_files:
      - split: train
        path: data/rat_positions.jsonl
  - config_name: rat_rollouts
    data_files:
      - split: train
        path: data/rat_rollouts.jsonl

rat

Rollouts: 920 rows written (920 successful after retries, 0 still failed). finish_reason counts: {'stop': 920}. Positions: 184 sentences over 2 instances, 0 refused. Importance: 182 rows, 182 scored, alpha=0.1, mean 0.0000.

Method: Thought Anchors resampling importance. For sentence i, the set is n_rollouts continuations from the trace prefix before s_i; importance is KL(P_i || P_{i+1}) over bucketed answers. Model: Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218), thinking on.

Parameters

{
  "sampling": {
    "temperature": 0.9,
    "top_p": 0.95,
    "top_k": 20,
    "min_p": 0.0,
    "max_tokens": 32768,
    "seed": "not sent"
  },
  "rollout_caps": {
    "max_tokens": 16384,
    "margin": 64,
    "min_room": 256
  },
  "n_rollouts": 5,
  "n_instances": 2,
  "max_sentences": 400
}

Sources and runs

Tagged traces: connections-dev/create-interpret-task-general-tagged @ 7a7f2b90ec22e5cdb5e9afce6fddf0bb4d9d4c12. Raw traces: connections-dev/create-interpret-task-general @ 4affe1fc0e49bbf908e169c827457c7a02d42fd8. Skipped instances (too many sentences): 0. Run commands:

/scratch/mw4141/code/create_interpret/anchors/cli.py rollout --tagged-repo connections-dev/create-interpret-task-general-tagged --samples-repo connections-dev/create-interpret-task-general --subset rat --n-instances 2 --max-sentences 400 --n-rollouts 5 --max-tokens 16384 --margin 64 --min-room 256 --concurrency 64 --out-dir /scratch/mw4141/anchors_out/smoke --push-to-hub --push-org connections-dev --push-repo-name create-interpret-task-general-anchors-smoke --push-visibility private --config configs/general_full.json

git sha(s): 9a6589e5467700aebd5edca6d77de4dc14ae5f68. Full manifest: manifests/rat.json.

rat_rollouts columns

Column Description
instance_id / sample_idx instance and the index of the tagged base trace the prefix was cut from
sentence_idx the sentence the prefix stops before (the model writes its own replacement for it)
rollout_idx 0..n_rollouts-1 within the set
content the continuation only (starts right after the prefix; thinking, , answer inline)
finish_reason / success / error vLLM finish_reason; false if the request failed
max_tokens / prompt_tokens per-request generation cap and prefix length in tokens
request_usage usage block of the whole n-sample request

rat_positions columns

Column Description
sentence_idx / sentence / tags position, its text and function tags from the tagged repo
n_sentences / rel_pos trace length and sentence_idx / (n_sentences - 1)
prefix_chars / prompt_tokens / max_tokens prefix length (chars, tokens) and the cap used
status / reason ok, or refused (not enough room in the context window) with the reason

rat_importance columns

Column Description
counts bucketed answer counts of the set at this sentence ( = no extractable answer)
n_ok / n_failed / n_no_answer completions used / failed requests / no extractable answer
top_answer most common bucket
importance KL(P_i
alpha smoothing used