Skip to content

SLM-300 + SLM-250: self-context exposure-bias curriculum + LOT1-01 not_authorized #895

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Tyler-R-Kendrick merged 4 commits into from
Jul 25, 2026
Merged

SLM-300 + SLM-250: self-context exposure-bias curriculum + LOT1-01 not_authorized #895

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
d99ce24
SLM-300: wire self-context exposure-bias curriculum (AP-015)
claude Jul 25, 2026
72f5de8
SLM-300: address CodeRabbit review on PR #895
claude Jul 25, 2026
2da5a2e
fix: restore main CI after #920 (levers, tokenizers, vocab remap)
Tyler-R-Kendrick Jul 25, 2026
d3b1c9c
fix(SLM-300): prepend ops.vocab no-bump after rebase onto main
Tyler-R-Kendrick Jul 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
333 changes: 333 additions & 0 deletions docs/design/iter-slm300-self-context-exposure-bias-20260725.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,333 @@
{
"arms": [
{
"arm_label": "A_control",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 0% seed 0",
"fixture-only: no model trained",
"policy origin mixture: gold=1.00 model_current=0.00 model_lagged=0.00",
"mixture zero reproduces legacy gold-only corruption exactly"
],
"policy_origin_mixture": {
"gold": 1.0,
"model_current": 0.0,
"model_lagged": 0.0
},
"run_id": "slm300_fixture_A_control_s0",
"seed": 0,
"self_context_rate": 0.0,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "A_control",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 0% seed 1",
"fixture-only: no model trained",
"policy origin mixture: gold=1.00 model_current=0.00 model_lagged=0.00",
"mixture zero reproduces legacy gold-only corruption exactly"
],
"policy_origin_mixture": {
"gold": 1.0,
"model_current": 0.0,
"model_lagged": 0.0
},
"run_id": "slm300_fixture_A_control_s1",
"seed": 1,
"self_context_rate": 0.0,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "A_control",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 0% seed 2",
"fixture-only: no model trained",
"policy origin mixture: gold=1.00 model_current=0.00 model_lagged=0.00",
"mixture zero reproduces legacy gold-only corruption exactly"
],
"policy_origin_mixture": {
"gold": 1.0,
"model_current": 0.0,
"model_lagged": 0.0
},
"run_id": "slm300_fixture_A_control_s2",
"seed": 2,
"self_context_rate": 0.0,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC10",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 10% seed 0",
"fixture-only: no model trained",
"policy origin mixture: gold=0.90 model_current=0.05 model_lagged=0.05"
],
"policy_origin_mixture": {
"gold": 0.9,
"model_current": 0.05,
"model_lagged": 0.05
},
"run_id": "slm300_fixture_SC10_s0",
"seed": 0,
"self_context_rate": 0.1,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC10",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 10% seed 1",
"fixture-only: no model trained",
"policy origin mixture: gold=0.90 model_current=0.05 model_lagged=0.05"
],
"policy_origin_mixture": {
"gold": 0.9,
"model_current": 0.05,
"model_lagged": 0.05
},
"run_id": "slm300_fixture_SC10_s1",
"seed": 1,
"self_context_rate": 0.1,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC10",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 10% seed 2",
"fixture-only: no model trained",
"policy origin mixture: gold=0.90 model_current=0.05 model_lagged=0.05"
],
"policy_origin_mixture": {
"gold": 0.9,
"model_current": 0.05,
"model_lagged": 0.05
},
"run_id": "slm300_fixture_SC10_s2",
"seed": 2,
"self_context_rate": 0.1,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC25",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 25% seed 0",
"fixture-only: no model trained",
"policy origin mixture: gold=0.75 model_current=0.12 model_lagged=0.12"
],
"policy_origin_mixture": {
"gold": 0.75,
"model_current": 0.125,
"model_lagged": 0.125
},
"run_id": "slm300_fixture_SC25_s0",
"seed": 0,
"self_context_rate": 0.25,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC25",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 25% seed 1",
"fixture-only: no model trained",
"policy origin mixture: gold=0.75 model_current=0.12 model_lagged=0.12"
],
"policy_origin_mixture": {
"gold": 0.75,
"model_current": 0.125,
"model_lagged": 0.125
},
"run_id": "slm300_fixture_SC25_s1",
"seed": 1,
"self_context_rate": 0.25,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC25",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 25% seed 2",
"fixture-only: no model trained",
"policy origin mixture: gold=0.75 model_current=0.12 model_lagged=0.12"
],
"policy_origin_mixture": {
"gold": 0.75,
"model_current": 0.125,
"model_lagged": 0.125
},
"run_id": "slm300_fixture_SC25_s2",
"seed": 2,
"self_context_rate": 0.25,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC50",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 50% seed 0",
"fixture-only: no model trained",
"policy origin mixture: gold=0.50 model_current=0.25 model_lagged=0.25"
],
"policy_origin_mixture": {
"gold": 0.5,
"model_current": 0.25,
"model_lagged": 0.25
},
"run_id": "slm300_fixture_SC50_s0",
"seed": 0,
"self_context_rate": 0.5,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC50",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 50% seed 1",
"fixture-only: no model trained",
"policy origin mixture: gold=0.50 model_current=0.25 model_lagged=0.25"
],
"policy_origin_mixture": {
"gold": 0.5,
"model_current": 0.25,
"model_lagged": 0.25
},
"run_id": "slm300_fixture_SC50_s1",
"seed": 1,
"self_context_rate": 0.5,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
},
{
"arm_label": "SC50",
"binder_reference_recovery": null,
"binding_aware_meaningful_v2_rate_strict": null,
"checkpoint_uri": null,
"denoising_round_recovery": null,
"notes": [
"planned self-context rate 50% seed 2",
"fixture-only: no model trained",
"policy origin mixture: gold=0.50 model_current=0.25 model_lagged=0.25"
],
"policy_origin_mixture": {
"gold": 0.5,
"model_current": 0.25,
"model_lagged": 0.25
},
"run_id": "slm300_fixture_SC50_s2",
"seed": 2,
"self_context_rate": 0.5,
"status": "fixture_planned",
"train_loss": null,
"validity_regression": null
}
],
"curriculum_id": "self-context-scheduled-corruption",
"manifest": {
"base_recipe": {
"batch_size": 4,
"checkpoint_sync": false,
"context_backend": "hf_local_files_only",
"device": "cpu",
"eval_version": "remediated",
"learning_rate": 0.0003,
"max_wall_minutes": 3.0,
"seed": 0,
"steps": 32,
"train_version": "e218_schema_normalized_judge_v5"
},
"base_recipe_hash": "11a333490bd3913a8ed1c573d36d1e6c12a38b000253611f5de34ecd021b8e0d",
"checkpoint_bucket": "hf://buckets/TKendrick/OpenUI",
"claim_class": "wiring",
"curriculum_id": "self-context-scheduled-corruption",
"falsifier": "Self-context mixture produces no paired strict-semantic recovery gain over the gold-only (rate=0) control, or any gain requires validity regression greater than 0.01 or more total tokens/updates than the matched control.",
"hypothesis": "Train-decode state mismatch (exposure bias) contributes to binding/reference recovery failures, and mixing verifier-aware, model-generated partial states into the corruption schedule at a moderate self-context rate improves recovery without validity regression or a hidden training-budget increase.",
"lagged_policy_share": 0.5,
"matrix_set": "self-context-exposure-bias",
"matrix_version": "ap015-v1",
"parent_checkpoint_uri": null,
"seeds": [
0,
1,
2
],
"self_context_rates": [
0.0,
0.1,
0.25,
0.5
],
"status": "not_run"
},
"matrix_set": "self-context-exposure-bias",
"matrix_version": "ap015-v1",
"run_id": "slm300_fixture",
"status": "fixture",
"version_stamp": {
"code_commit": "caa144b0524a758dbdefa3b8e70b563dc126e11a",
"code_dirty": true,
"components": {
"harness.experiments": "v106"
},
"stamp_schema": "version_stamp/v1",
"stamped_at": "2026-07-25T08:43:03.262457+00:00"
}
}
Loading
Loading