mrsndmn/progressive_cramming_trajectories

Progressive Cramming Trajectories

Code & method: github.com/FusionBrainLab/progressive_cramming — reference implementation of progressive cramming (the public repo).

Per-stage progressive cramming trajectories for four pretrained LLM families. This dataset backs the tab:progressive_modifications table of the paper: it records, for each document, the full sequence of converged soft-prompt ("memory") embeddings produced as the compressed context is progressively grown one stage at a time.

Each row is one converged cramming stage — a single point on a document's optimization trajectory. To reconstruct a document's trajectory, take all rows with the same sample_id and sort by (stage_index, stage_seq_len).

Source documents: 50 held-out PG19 passages (pg19_1k), sequence length 4096.

Layout

One config per model family, two splits per config:

Config (config_name)	baseline arm	lowdim arm
Llama-3.1-8B	full-dim per-sample embedding, lr=0.1	low-dim projection, dim=256, lr=0.1
pythia-1.4b	full-dim per-sample embedding, lr=0.5	low-dim projection, dim=256, lr=0.5
SmolLM2-1.7B	full-dim per-sample embedding, lr=0.1	low-dim projection, dim=256, lr=0.1
gemma-3-4b-pt	full-dim per-sample embedding, lr=0.1	low-dim projection, dim=32, lr=0.1

• baseline — the embedding is optimized directly in the model's full hidden space.
• lowdim — the embedding is parametrized by a low-dimensional projection (over-parametrised per-sample embedding), which yields more compressed tokens per document and a much longer trajectory.

Usage

from datasets import load_dataset

# one model family, one arm
ds = load_dataset("mrsndmn/progressive_cramming_trajectories", "SmolLM2-1.7B", split="baseline")

# reconstruct sample 0's trajectory (ordered list of converged stages)
traj = ds.filter(lambda r: r["sample_id"] == 0).sort(["stage_index", "stage_seq_len"])

The embedding columns are float64 lists of shape [num_compression_tokens, hidden_size] and are large; use .select_columns([...]) / streaming to avoid materialising them all.

Columns

Column	Meaning
sample_id	Document index (0–49).
stage_index	Progressive stage number (trajectory step).
stage_seq_len	Number of source tokens crammed at this stage.
text	The crammed text.
embedding	Converged compression-token embedding at this stage — shape [num_compression_tokens, hidden_size]. The trajectory point.
orig_embedding	Back-compat duplicate of embedding (kept identical for legacy eval scripts).
pca_coefficients_to_save	Low-dim coefficients in newer runs; null in these runs.
initialization_embedding	Embedding used to initialize this stage's optimization.
final_loss	Final cross-entropy loss at convergence for this stage.
final_convergence	Final token-recovery accuracy at convergence.
num_input_tokens	Number of input tokens for the stage.
num_compression_tokens	Number of compression ("memory") tokens.
hidden_size	Model hidden dimension.
loss_type, dtype, model_checkpoint	Run metadata.
max_optimization_steps_per_sample, convergence_threshold	Optimization config.
steps_taken	Optimization steps taken to converge this stage.
information_gain_bits	Information gain (bits) attributed to this stage.

Known caveat: saved-embedding off-by-one

These artifacts were generated before the ProgressiveCrammingTrainer off-by-one fix. The trainer stepped the optimizer once more after the converged forward pass, so the saved embedding (and orig_embedding, information_gain_bits) are one optimization step past the converged state E_t whose final_loss / final_convergence were recorded. When reconstructing prefixes from embedding, expect occasional large-margin mismatches against the reported final_convergence (≈1.0). The scalar metrics (final_loss, final_convergence, steps_taken) are unaffected.

Provenance

Generated by progressive-cramming runs and published from the project's research codebase. A lean, self-contained reference implementation of the method (full / progressive / low-dim cramming) is in the public repository: github.com/FusionBrainLab/progressive_cramming. The arm/run definitions mirror the paper's progressive_modifications table.