Addressed your concerns, I got these results:
```json
{
“experiment_id”: “litm_with_baselines”,
“date”: “2026-07-28T08:54:24.327926”,
“model_details”: {
“base_model”: “unsloth/mistral-7b-instruct-v0.3-bnb-4bit”,
“quantization”: “4-bit NF4”
},
“scorer_details”: {
“checkpoint_path”: “/mnt/juegos/proyectos/especiales/token-importance/checkpoints/v8_hard_anchor_final”,
“scorer_type”: “query_trained_head”,
“architecture”: “ImportanceUpdateHead (cross-attention + RMSNorm + linear)”,
“training_objective”: “ERT (KL divergence, token eviction)”,
“frozen”: true,
“task_specific_retraining”: false,
“training_data”: “MS-MARCO train split”
},
“dataset”: {
“source”: “MS-MARCO passage QA (train split)”,
“num_examples”: 60,
“num_positions”: 3,
“position_bucket_definition”: “early (index 0), middle (k//2), end (k-1)”,
“examples_are_paired”: false,
“independently_generated_per_position”: true,
“passages_per_context”: 5,
“seed”: 42
},
“pipelines”: [
“baseline”,
“lexical (TF-IDF)”,
“TIS”,
“oracle (gold-first)”
],
“command_used”: “scripts/run_litm_with_baselines.py --n-examples 60 --seed 42”
}
```
And this result statement too:
The rest of your concerns have been addressed and will be incorporated to the final results, to be published here soon, in a new post! Stay tuned!