agurung/colar-qwen25-7b-ff-post-sft

CoLAR Qwen2.5-7B Flawed Fictions Post-SFT

This repository contains a historical Qwen2.5-7B CoLAR checkpoint exported to a Hugging Face-compatible layout. The repo root works for standard decoding (Transformers/vLLM), and extra_state.pt preserves the CoLAR latent head for latent decoding.

What Is In This Repo

• Standard HF model files at repo root (config.json, model*.safetensors, tokenizer files)
• extra_state.pt with the latent policy and embedding weights required for CoLAR latent decoding
• export_meta.json from the local export step
• latent_metadata.json with structured provenance and hyperparameters
• artifacts/source_wandb_config.yaml, artifacts/source_wandb_summary.json, and artifacts/source_wandb_metadata.json
• artifacts/source_all_config.json parsed from the original WandB all_config

Source Provenance

Field	Value
WandB run	owf320j4
Run date	2025-10-27
Task	Flawed Fictions continuity error detection
Stage	post-SFT
Base model	Qwen/Qwen2.5-7B-Instruct
Git commit	e385c665547298afafef9cffb04d1e530f8e57c4
Host / GPU	alexgurung-fftest-s8rvj-66w4n / NVIDIA H200
Original checkpoint	/mnt/volume3/baseline_colar/colar/oldlogs/colar/colar-experiments/owf320j4/checkpoints/pt_best_ckpt
Generalization slug	colar_ff_post_sft

Usage

Standard decoding

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_dir = "agurung/colar-qwen25-7b-ff-post-sft"
model = AutoModelForCausalLM.from_pretrained(repo_dir, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(repo_dir)

Latent decoding

from huggingface_hub import snapshot_download
from pathlib import Path

local_dir = Path(snapshot_download("agurung/colar-qwen25-7b-ff-post-sft"))
llm_dir = local_dir
extra_state = local_dir / "extra_state.pt"

Example evaluator invocation:

python -m litereason.experiments.latent_eval.colar.evaluate_colar   --llm-dir "$LOCAL_DIR"   --extra-state "$LOCAL_DIR/extra_state.pt"   --test-file litereason/experiments/generalization/data/gsm8k.jsonl   --prompt-variant standard   --save-preds preds_gsm8k_standard.jsonl   --num-samples 5   --use-chat-template

Key Hyperparameters

Field	Value
Trainer strategy	deepspeed_stage_2_offload
Max epochs	5
Precision	bf16
Global batch size	1
Validation batch size	4
Compression factor	5
Max latent forward	64
Latent temperature	1.0
Group size	8
Experience batch size	1
Train samples / epoch	64

Reported Metrics

Metric	Value
monitor	8.439671516418457
val/reward	-8.439671516418457
val/loss	8.439671516418457
val/ce_loss	5.265884876251221
val/embed_modeling_loss	3.1737866401672363
epoch	4
trainer/global_step	75

Notes

• Historical CoLAR Flawed Fictions supervised fine-tuning checkpoint.
• This export preserves the latent head in extra_state.pt for latent decoding.

Local Reference Paths

• WandB config: /mnt/volume3/baseline_colar/colar/oldlogs/colar/wandb/run-20251027_171940-owf320j4/files/config.yaml
• WandB summary: /mnt/volume3/baseline_colar/colar/oldlogs/colar/wandb/run-20251027_171940-owf320j4/files/wandb-summary.json
• WandB metadata: /mnt/volume3/baseline_colar/colar/oldlogs/colar/wandb/run-20251027_171940-owf320j4/files/wandb-metadata.json
• Local export dir: /mnt/disk/baseline_colar/colar/exported_hf/owf320j4